GDDR5X Puts Up a Fight Against HBM, AMD and NVIDIA Mulling Implementations

[lilhasselhoffer]

...
And therefore my argument is the opposite of silly. First try to understand before you run out and call anyone "silly". You are very quick to judge or harass persons, not the first time you've done that - you do it way too frequently, and shouldn't do it at all.

I only see HBM1/2 for highend cards now or in 2016 because its too expensive for lower cards - therefore it's (GDDR5X) good for middle to semi highend cards I'd say. I even would bet on that. You really expect a premier technology to be used on middle to semi high end cards? Me not. GDDR5X has a nice gap there to fill, I think.

Well that's no problem. If it arrives with HBM2, they can plan for it and produce cards with it (GDDR5X). I don't see why this would be a problem. And as I said, I only see HBM2 on highend or highest end (650$) cards. I think the 400 (if any)/500/650$ cards will have it - so everything cheaper than that will be GDDR5X, and that's not only "cheap cards". I don't see 200-400 as "cheap". Really cheap is 150 or less and this is GDDR5-land (not even GDDR5X) I think.

I contest these three points. Everything else I may not agree with, but there isn't any reason to believe it isn't possible.


Here's the post I quoted, paring out the unnecessary bits. I've highlighted the silliness. If I weren't being generous, I'd call garbage statments which are either factually incorrect or useless.

...I didn't talk about HBM one bit. But you're still wrong, HBM has already proven itself on Fiji cards - it's a fact that they would be even more bandwidth limited without HBM, as overclocking HBM increased performance of every Fiji card further. This has proven that Fiji has not too much bandwidth, it has proven that it can't have enough bandwidth. Besides HBM made Fiji possible - there would be no Fiji without HBM. The same chip with GDDR5 would've taken more than 300 W TDP, a no-go, or would have needed a lower clock speed, which is a no-go too, because AMD wanted to achieve Titan X-like performance. The 275W TDP of Fury X was only possible with HBM on it. So not only DID the increased bandwidth help (a lot), it helped making the whole card possible at all.

As you can see, most of this statement is silly. Allow me to tear into it though.
1) it's a fact that they would be even more bandwidth limited without HBM,
Really? Is it a fact? How then does Nvidia have the performance it has with Titan? You make a statement of false equivocation, based upon a faulty premise. Because HBM is designed to have higher bandwidth it must therefore be responsible for performance. Where are your facts?
2) overclocking HBM increased performance of every Fiji card further
Again, facts. What I've seen is overclocking HBM leads to 50% or less returns on improvement. For example, the cited 8% overclock only returns 4% increased numerical results. Technically overclocking does increase performance, but you're weaseling out of this argument by saying any increase is an increase. When 50% of your added effort is wasted without seeing real improvement then it isn't really a reasonable improvement.
3) Fiji has not too much bandwidth
This goes back to point two. If Fury X (not Fiji in general) was bandwidth limited an 8% increase in clocks would yield somewhere near 8% of improved performance. It doesn't, therefore your point is invalid. Rather than dismissing it, I call it a silly and unsubstantiated point.
4) proven that it can't have enough bandwidth
Same as 3.
5) there would be no Fiji without HBM
Except you're wrong. There would be no Fury inside of the form factor and thermal envelope they chose, that doesn't mean Fiji wouldn't exist. You're equating two entirely separate and unrelated topics without factual basis here.
6) 275W TDP of Fury X
Artificially chosen value by AMD. This is irrelevant to the implementation of HBM (as demonstrated by Nvidia).
7) So not only DID the increased bandwidth help (a lot), it helped making the whole card possible at all
You reiterated all of your previous points in a single sentence. Congratulations, but a house built without a solid foundation is going to collapse rather easily. You've drawn all of these conclusions, in the face of existing data that proves you wrong. My teachers called it cute when I did this in school. My bosses fired me for incompetence. You are denying reality, and therefore are either an idiot or silly. I choose to give you the benefit of doubt and assume silliness. We've all been guilty of that at some point in time.



As to HBM2 not being available/cost effective, may I ask what exactly you expect of GDDR5X? It's an as yet unmanufactured standard, without substantial testing behind it. It may be largely plug and play with older controllers, but it still has to be made by somebody. This means added costs as the process is proven out, additional costs for redesigns of controllers to actually see the benefits of GDDR5X (why would you switch to what has to be a more expensive type of memory while cheaper stuff is more readily available), and supply issues all their own.

What you're arguing for is that in the midst of pushing out HBM, both AMD and Nvidia will push out another new standard. Why? Why would you ever completely retool everything, with less than 12 months to design, test, rework, prototype, and have manufacturing specifications for a product line? It'd be insane to do so.

Let's offer the benefit of doubt again, and acquiesce to your theory (based on nothing) that 90% of cards will not be HBM based. In order to accept that we have to make the assumption that HBM2 is not being produced right now, and will in fact only see production late next year. Why would AMD and Nvidia let that happen? They know that their new processes will finally be out next year. They know that the shrink will produce more performance gains than the last two redesigns, because of its huge magnitude. They know that Pascal and Arctic Islands will be the time that everybody re-evaluates 3-4 year old cards and decides it's time to evaluate an upgrade. Knowing all of this, how do you come to the conclusion that they aren't already starting on HBM2 chip orders (yes, it was the interposer that was the issue with HBM1, I know)?


I refer to your statements as silly, because they make no logical sense. I don't directly call your opinions idiotic, because you've demonstrated a grasp on reality. Our opinions may differ, but you deserve the respect of being proven wrong rather than being dismissed out of hand for for saying things that are unconnected to observable and demonstrable reality. It would help your argument to bring facts to the table though. It's hard to argue point like "overclocking makes things faster" is inaccurate, when you don't go out and at least try to have factual support for your statements.

As to your personal opinions of me, I don't care. You are more than welcome to call me an ass, and there's no reason to defend myself against it. In the last month I've been misinterpreted so as to call everyone from West Virginia as victims of a lack of genetic diversity (further being stretched to personal insultation of a person I don't know), I've been accused of calling southerners all hicks and rednecks (despite never using the term), and I've more than once been proven wrong. None of us, in the US at least, have the right to not be offended. The point of a forum is to raise a substantive argument based upon facts, and have poorer arguments torn apart by reality. It's the only way we can make sure our beliefs are rooted in reality, and not fantasy.

 

[HumanSmoke]

5) there would be no Fiji without HBM
Except you're wrong. There would be no Fury inside of the form factor and thermal envelope they chose, that doesn't mean Fiji wouldn't exist. You're equating two entirely separate and unrelated topics without factual basis here.

Kanen is closer to being right than wrong on this point. Technically you could claim that AMD could apply the name Fiji to a rebadged Pitcairn if it so desired, but the fact of the matter is that AMD managed to produce a 4096 core GPU that is a 45.5% gain over Hawaii for a 36% increase in die size - well beyond any compensations you could make for fixed sized logic ( command processor, VCE and UVD engines, and PCI-E I/O etc) bearing in mind AMD also increased L2 cache by 100% with Fiji.
As it is, at 596mm² is still very close to TSMC's maximum die limit (625mm²). The simple fact of the matter is AMD would have had to make some substantial compromises with Fiji if it had required a GDDR5 interface. They still could have called it Fiji, but it wouldn't have been the same Fiji. Depends on whether you're arguing technical specification and performance, or semantics.

 

[lilhasselhoffer]

Kanen is closer to being right than wrong on this point. Technically you could claim that AMD could apply the name Fiji to a rebadged Pitcairn if it so desired, but the fact of the matter is that AMD managed to produce a 4096 core GPU that is a 45.5% gain over Hawaii for a 36% increase in die size - well beyond any compensations you could make for fixed sized logic ( command processor, VCE and UVD engines, and PCI-E I/O etc) bearing in mind AMD also increased L2 cache by 100% with Fiji.
As it is, at 596mm² is still very close to TSMC's maximum die limit (625mm²). The simple fact of the matter is AMD would have had to make some substantial compromises with Fiji if it had required a GDDR5 interface. They still could have called it Fiji, but it wouldn't have been the same Fiji. Depends on whether you're arguing technical specification and performance, or semantics.

Maybe my understanding is wrong here then, so help me see where you're coming from.

In a traditional GPU the core is at the center, and the memory controller is functionally going to be the outside of the die. Thus, the limits of the silicon are further shrunk down by having the memory controller eat into them. With HBM the limit is moved from the die, to the interposer. Because the interposer is what is interfacing the GPU to its memory you run into the exact same spacing constraints, further complicated by having to connect the HBM modules to not only the interposer, but to one another in order to expand capacity (read: your earlier comments about the micro-bumps).

This effectively changes the limitations from one element to another, without ever removing the limitations. It also adds another complex manufacturing element to the mix, as the interposer was the reason Fury X was supposedly in small supply (with their small form factor offering basically being non-existent).


Tell me how this changes anything? I'm not aware of AMD suddenly having the ability to produce gigantic interposers (if you can prove me wrong that would be great, and I'd be happy to apologize for my ignorance). It could therefore safely be stated that the size restrictions are effectively unchanged. You have to contend with more heat, but again you can stretch chip placement out when they aren't forced to all fit onto an interposer. Put simply, Nvidia demonstrated the most elegant counter argument to Fury's performance by beating it roundly at the same or a cheaper price. While it makes me feel dirty to say it, Nvidiots have a point here. The Red team has yet to show us innovation, despite elegantly demonstrating that HBM isn't another Bulldozer.

I need a shower now. Saying Nvidia is reasonable and calling out Bulldozer just makes me feel unclean.

 

[Casecutter]

chip has a lifetime of 6-8 months

I just meant that all AMD needed was for Hawaii to hold its ground from it's rebrand start date of July 2015 to when 14nm comes along.

 

[HumanSmoke]

Maybe my understanding is wrong here then, so help me see where you're coming from.

In a traditional GPU the core is at the center, and the memory controller is functionally going to be the outside of the die. Thus, the limits of the silicon are further shrunk down by having the memory controller eat into them.

Correct. Memory I/O is obviously arrayed around the periphery of the GPU since memory chip -to- IMC need to be of equal distance (see the Tahiti die shot I posted in post #19)

With HBM the limit is moved from the die, to the interposer. Because the interposer is what is interfacing the GPU to its memory you run into the exact same spacing constraints, further complicated by having to connect the HBM modules to not only the interposer, but to one another in order to expand capacity (read: your earlier comments about the micro-bumps).

Only partially correct. Fiji still retains (simpler) IMC's and PHY, as AMD's own block diagrams from the marketing slide deck show.

This effectively changes the limitations from one element to another, without ever removing the limitations.

The interposer is not the GPU. The discussion centres around you contesting that Fiji would exist as it currently is (resource wise) even if saddled with the need for GDDR5 support.

It also adds another complex manufacturing element to the mix, as the interposer was the reason Fury X was supposedly in small supply (with their small form factor offering basically being non-existent).

The interposer is not the issue with Fury X's supply constraint - it is the micro-bumping assembly and verification (both of the HBM die stacks and HBM -> substrate). Tooling for the latter has only recently on line for packagers like Amkor (who assemble the Fury+HBM+ interposer package) for both 2.5D and the next gen of 3D stacked products.

Tell me how this changes anything? I'm not aware of AMD suddenly having the ability to produce gigantic interposers (if you can prove me wrong that would be great, and I'd be happy to apologize for my ignorance). It could therefore safely be stated that the size restrictions are effectively unchanged.

Interposer size isn't the overriding constraint. The Fiji/HBM package is actually larger than the interposer is sits upon (and the metal layer that separates them). The interposer just needs to be large enough to have contact for the required micro-bumps on the underside of the HBM stacks. Making the GPU+HBM package larger (or the interposer smaller) just affects the density of the TSV's within the interposer substrate. Since Fiji combined a lot of relatively new ground for AMD (their largest GPU by a handsome margin and the first HBM implementation), they decided upon a mature/conservative (read old) 65µm process node, rather than add another layer of potential issue. Physical manufacturing size constraints determined Fiji. The interposer tech determines the chances of commercial success ( yield/delivery and manufacturing cost).

I just meant that all AMD needed was for Hawaii to hold its ground from it's rebrand start date of July 2015 to when 14nm comes along.

I wouldn't expect GDDR5X to arrive appreciably before the next generation of cards. I doubt any current cards would be retrofitted with GDDR5X. If 14nm (or more likely 16nm) is delayed and a refresh of lineup is "required", maybe you'll see implementation - but I doubt it. TSMC's 16nmFF+ seems to be on track at the moment and any potential refresh would have a very short linespan indeed. Also bear in mind that AMD no longer designs or lays out it own I/O (including memory) logic - that is now in the hands of Synopsys.

 

[lilhasselhoffer]

Correct. Memory I/O is obviously arrayed around the periphery of the GPU since memory chip -to- IMC need to be of equal distance (see the Tahiti die shot I posted in post #19)

Only partially correct. Fiji still retains (simpler) IMC's and PHY, as AMD's own block diagrams from the marketing slide deck show.

The interposer is not the GPU. The discussion centres around you contesting that Fiji would exist as it currently is (resource wise) even if saddled with the need for GDDR5 support.

The interposer is not the issue with Fury X's supply constraint - it is the micro-bumping assembly and verification (both of the HBM die stacks and HBM -> substrate). Tooling for the latter has only recently on line for packagers like Amkor (who assemble the Fury+HBM+ interposer package) for both 2.5D and the next gen of 3D stacked products.

Interposer size isn't the overriding constraint. The Fiji/HBM package is actually larger than the interposer is sits upon (and the metal layer that separates them). The interposer just needs to be large enough to have contact for the required micro-bumps on the underside of the HBM stacks. Making the GPU+HBM package larger (or the interposer smaller) just affects the density of the TSV's within the interposer substrate. Since Fiji combined a lot of relatively new ground for AMD (their largest GPU by a handsome margin and the first HBM implementation), they decided upon a mature/conservative (read old) 65µm process node, rather than add another layer of potential issue. Physical manufacturing size constraints determined Fiji. The interposer tech determines the chances of commercial success ( yield/delivery and manufacturing cost).

I wouldn't expect GDDR5X to arrive appreciably before the next generation of cards. I doubt any current cards would be retrofitted with GDDR5X. If 14nm (or more likely 16nm) is delayed and a refresh of lineup is "required", maybe you'll see implementation - but I doubt it. TSMC's 16nmFF+ seems to be on track at the moment and any potential refresh would have a very short linespan indeed. Also bear in mind that AMD no longer designs or lays out it own I/O (including memory) logic - that is now in the hands of Synopsys.

Then I have a few things to say.

First, my understanding was flawed. Thank you for rectifying it.

Second, by a very narrow definition Fiji couldn't exist if HBM was not available. That definition requires that the die could not be expanded at all, that the interposer could take a GPU as big as the dies could be manufactured, and that there was no more efficient organization in the silicon which could have improved performance. I acquiesce that instead of being technically precise, and stating it as such, I didn't give a technically correct answer.

Third, the interposer seems to still be a size restraint here. I can understand why AMD wanted to save money and go with 65 nm (just like Intel having its PCHs generations behind the CPU to save money), but theoretically not having a limit is different than the practical limits which AMD set. I'll give you the point though, because this argument is more philosophical than based in hard numbers.

 

[HumanSmoke]

Third, the interposer seems to still be a size restraint here. I can understand why AMD wanted to save money and go with 65 nm (just like Intel having its PCHs generations behind the CPU to save money), but theoretically not having a limit is different than the practical limits which AMD set. I'll give you the point though, because this argument is more philosophical than based in hard numbers.

Interposer's that fit AMD's requirement were probably the limiting factor. The common number associated with interposers is $1-2 per/mm² which works out pretty expensive once you start to work out the math, and a prime reason why interposers have been the province of expensive FPGA's like Xilinx's Virtex-7 (which sits upon an interposer of 775mm² incidentally). Interposers can be as large as you like since they aren't bound by lithography reticule limits (only the interconnects are reticle bound - hence the copper interconnect area not extending across the whole of the interposer. The non-interconnect area appearing green on this Fiji package picture), which is why wafer size will ultimately determine the absolute size of the interposer (such as TSMC's plans - who incidentally provide the interposer for the aforementioned Xilinx Virtex-7.

 

[Kanan]

As you can see, most of this statement is silly. Allow me to tear into it though.
1) it's a fact that they would be even more bandwidth limited without HBM,
Really? Is it a fact? How then does Nvidia have the performance it has with Titan? You make a statement of false equivocation, based upon a faulty premise. Because HBM is designed to have higher bandwidth it must therefore be responsible for performance. Where are your facts?

Silly is that you compare Nvidia cards with AMD cards and doesn't understand their differences one bit. As already said (and ignored by you), AMD cards don't have the nice compression Maxwell has, so it needs 512 bit with high MHz or HBM - Fiji needs HBM, GDDR5 wouldn't suffice I think, because even 1750 MHz 512 bit is way below 512 GB/s bandwidth, what Fiji now has with HBM and still scaling with it if you overclock it, what essentially proves that HBM is worth it, even just for the extra bandwidth it delivers, set aside the power reduce which made Fiji/Fury X possible at all.

2) overclocking HBM increased performance of every Fiji card further
Again, facts. What I've seen is overclocking HBM leads to 50% or less returns on improvement. For example, the cited 8% overclock only returns 4% increased numerical results. Technically overclocking does increase performance, but you're weaseling out of this argument by saying any increase is an increase. When 50% of your added effort is wasted without seeing real improvement then it isn't really a reasonable improvement.

It is a improvement, every bit helps. Nothing to do with "weaseling out", rofl, you simply don't understand my point, thats all. Again: HBM has high bandwidth, Fiji needs it, and overclocking HBM makes it even better - so what you said there on Page 1, about HBM being useless, is utter bullshit. And my overclocking-argument was just to prove how wrong your statement is, has nothing to do with scaling, what I know myself of, isn't the best, when overclocking HBM - but still it helps. And it is very reasonable, because you don't need more voltage for that +50-100 MHz on HBM, but you get free additional performance.

3) Fiji has not too much bandwidth
This goes back to point two. If Fury X (not Fiji in general) was bandwidth limited an 8% increase in clocks would yield somewhere near 8% of improved performance. It doesn't, therefore your point is invalid. Rather than dismissing it, I call it a silly and unsubstantiated point.

No your point is invalid. Just because it doesn't scale linearly doesn't mean its not bandwidth starved. AMD themself always speak of "how bandwidth limited cards are these days" and "always need more" - so AMD is wrong? Nope. Diminishing returns doesn't mean it is not bandwidth starved. Your point is just an unproven opinion of yours, nothing more.

4) proven that it can't have enough bandwidth
Same as 3.

Wrong.

5) there would be no Fiji without HBM
Except you're wrong. There would be no Fury inside of the form factor and thermal envelope they chose, that doesn't mean Fiji wouldn't exist. You're equating two entirely separate and unrelated topics without factual basis here.

Nope I'm very likely right with this. Without a possible Fury X, or say, a card that is strong enough to go face to face with the Titan X or 980 Ti, AMD would have scraped the entire line. Just "Nano" isn't worth it to do the Fiji+HBM+Interposer etc. thing, because Nano is not a very good selling card at all. Fury X is the first card - everything else is just a alternation to it, that's a 99,9999% fact, so if you delete Fury X because it's not possible without HBM, because TDP would be too high (this is a fact), you delete the entire Fury / Nano line. I thought that at this point everyone understood this - was written in articles everywhere, how important HBM was to design the Fiji/Fury-line of cards. Here on TPU, too, I think.

6) 275W TDP of Fury X
Artificially chosen value by AMD. This is irrelevant to the implementation of HBM (as demonstrated by Nvidia).

What? I just see a lot of senseless words here. Probably you didn't understand my point (again) and answered with some weird silly stuff. Plus, HBM is AMD tech, not Nvidia, you're completely off track here. HBM1 is a AMD exclusive, if AMD didn't want to make Fiji, HBM1 would not exist now - they would have skipped it for HBM2. But again - I don't think you understood my point. My point was, a Fury X with 275W TDP was only possible with HBM, not with GDDR5. Simple.

7) So not only DID the increased bandwidth help (a lot), it helped making the whole card possible at all
You reiterated all of your previous points in a single sentence. Congratulations, but a house built without a solid foundation is going to collapse rather easily. You've drawn all of these conclusions, in the face of existing data that proves you wrong. My teachers called it cute when I did this in school. My bosses fired me for incompetence. You are denying reality, and therefore are either an idiot or silly. I choose to give you the benefit of doubt and assume silliness. We've all been guilty of that at some point in time.

Yep, and I call your bluff or to say it in german words - you fell out of the window, you opened too wide and too fast. You're completely wrong on all or almost all points, certainly the most important ones about HBM and Fiji/Fury X.
Your psycho games are just utter childish bullshit, sorry. And I could laugh my ass off of you trying to insult me or call me a idiot. The idiot here is you, and it's not the first time I see you pissing off people or bullying around, or just being senselessly arrogant about just everything. And your other weakness is, that you just take about everything personal - THIS is silly. You're like that "Sony" guy, just somewhat alternated. Fact is, your pissing people off, because you're unable to debate in a civilized / friendly manner. Continue to do this and you're the first person on my black list, congratulations weirdo.

As to HBM2 not being available/cost effective, may I ask what exactly you expect of GDDR5X? It's an as yet unmanufactured standard, without substantial testing behind it. It may be largely plug and play with older controllers, but it still has to be made by somebody. This means added costs as the process is proven out, additional costs for redesigns of controllers to actually see the benefits of GDDR5X (why would you switch to what has to be a more expensive type of memory while cheaper stuff is more readily available), and supply issues all their own.

As I've understood it, GDDR5X is just a small evolution of GDDR5 and is easy to implement, saying, it does maybe not need new controllers etc. but I'm not sure about this, so take my answer with a grain of salt. I'm only sure HBM/HBM2 is expensive, because of the way it needs to be connected on the interposer and many companys working together to implement it.

What you're arguing for is that in the midst of pushing out HBM, both AMD and Nvidia will push out another new standard. Why? Why would you ever completely retool everything, with less than 12 months to design, test, rework, prototype, and have manufacturing specifications for a product line? It'd be insane to do so.

GDDR5X is just a small evolution, not a revolution so I don't see the problem here. HBM/HBM2 is the big story - and will in 2016 still be. But AMD has that already behind them at least partially.

Let's offer the benefit of doubt again, and acquiesce to your theory (based on nothing) that 90% of cards will not be HBM based. In order to accept that we have to make the assumption that HBM2 is not being produced right now, and will in fact only see production late next year. Why would AMD and Nvidia let that happen? They know that their new processes will finally be out next year. They know that the shrink will produce more performance gains than the last two redesigns, because of its huge magnitude. They know that Pascal and Arctic Islands will be the time that everybody re-evaluates 3-4 year old cards and decides it's time to evaluate an upgrade. Knowing all of this, how do you come to the conclusion that they aren't already starting on HBM2 chip orders (yes, it was the interposer that was the issue with HBM1, I know)?

I never said "90%" of cards. I said only (semi-)highend cards will have HBM2. That's all. And I didn't say "they aren't starting on HBM2 chip orders", thats just a wrong assumption you made.

I refer to your statements as silly, because they make no logical sense. I don't directly call your opinions idiotic, because you've demonstrated a grasp on reality. Our opinions may differ, but you deserve the respect of being proven wrong rather than being dismissed out of hand for for saying things that are unconnected to observable and demonstrable reality. It would help your argument to bring facts to the table though. It's hard to argue point like "overclocking makes things faster" is inaccurate, when you don't go out and at least try to have factual support for your statements.

I don't care what you think of me. You're certainly the most arrogant person I've seen in my entire life and I'm on the internet since I was 13 and I'm 29 years old, so it means something you're winning that "prize". As I already stated, if you continue your misguided behaviour towards me, you're the first person on my blacklist and I won't miss you for one second. And your arguments are already proven as wrong or misguided, so there's no point in what you try to do here.

As to your personal opinions of me, I don't care. You are more than welcome to call me an ass, and there's no reason to defend myself against it. In the last month I've been misinterpreted so as to call everyone from West Virginia as victims of a lack of genetic diversity (further being stretched to personal insultation of a person I don't know), I've been accused of calling southerners all hicks and rednecks (despite never using the term), and I've more than once been proven wrong. None of us, in the US at least, have the right to not be offended. The point of a forum is to raise a substantive argument based upon facts, and have poorer arguments torn apart by reality. It's the only way we can make sure our beliefs are rooted in reality, and not fantasy.

Well, I don't know what your parents did to you, but you seem to had a hard childhood. Life is about joy, not the hatred you are practicing. Well maybe not hatred but clearly in the direction of it. I would rather call it "disliking behaviour" or something of that sort. So you're an ass and you're happy with it and you try to make it legal by thinking its just the best way to be. So heres the surprise: it is NOT. Your behaviour is stressing everyone, a lot of your arrogance is based on you thinking you're extremely smart, but almost everything you said is extremely replacable or simply wrong. So continue your senseless ignorance - you are still wrong with your behaviour, and you can argue about that your entire life, you still will be wrong. Your behaviour is just bad. I tried to be tolerant towards you, when I first saw you quoted me and started the discussion with me, but I already knew you're difficult. But now you reached a point where my tolerance ends, I don't need random persons in the internet calling me a idiot or silly (or both), and then after it acting as it would be "justified". There are billion times more intelligent persons wandering the earth than you, and they are still humble and friendly. Think of that for a second if you CAN. You're just pitiful, sorry.

 

[Aquinus]

I understand this entire debate and I too could be missing something here. If GDDR5x is capable of doing up to 12GT/s, what is stopping AMD from going back to a 256-bit bus to reduce the size of the IMC to accommodate more compute? Doubling transfer rate and halving the size of the data bus gives you the same theoretical bandwidth. So I think we should keep in mind that faster memory doesn't necessarily mean that it needs to have a really wide bus.

This is starting to feel a whole lot like the serial versus parallel connection argument. Which is better: slow and wide or fast and narrow? In reality, it's probably somewhere in the middle.

 

[HumanSmoke]

I understand this entire debate and I too could be missing something here. If GDDR5x is capable of doing up to 12GT/s, what is stopping AMD from going back to a 256-bit bus to reduce the size of the IMC to accommodate more compute? Doubling transfer rate and halving the size of the data bus gives you the same theoretical bandwidth. So I think we should keep in mind that faster memory doesn't necessarily mean that it needs to have a really wide bus.

This is starting to feel a whole lot like the serial versus parallel connection argument. Which is better: slow and wide or fast and narrow? In reality, it's probably somewhere in the middle.

I think that is the crux of market segmentation.
High end/enthusiast cards (that pull double duty as pro parts) would benefit from HBM which will have both bus width and density advantages, but come at a BoM and manufacturing cost. Costs that can be recouped by the high pricing of the cards.
Anything from the lower performance - mainstream - entry level cannot absorb the additional cost that HBM implementation requires, and nor are those segments particularly bandwidth sensitive enough to make HBM a critical choice. Bearing in mind that die size, as you say, is either minimized. or the die area repurposed (compute/larger cache/more cores) to increase the effectiveness of the GPUs, vendors have a choice of reducing BoM or raising the performance bar. Micron has already announced that they begin production of 8Gbit (1GB) memory chips which effectively doubles the current density - giving the vendor choice of halving the number of chips and bus width but losing nothing in bandwidth or vRAM frame buffer...or choice of doubling one, or both. Again, it allows vendors to reduce BoM ( less chips, smaller PCB) or use the doubled density as a marketing tool - a 2, 3, or 4GB card suddenly becomes a 4, 6, or 8 GB card.
From my perspective, GDDR5X gives vendors better flexibility until HBM and other 2.5D/3D memory implementations become economically viable to use down the entire product stack.

FWIW, Samsung already sell 8Gbit GDDR5 chips, but because they have had the market to themselves, the cost is high (only the FirePro W9100 uses them as far as I'm aware - both for density and to shave a few watts off the power demand)

 

[lilhasselhoffer]

Silly is that you compare Nvidia cards with AMD cards and doesn't understand their differences one bit. As already said (and ignored by you), AMD cards don't have the nice compression Maxwell has, so it needs 512 bit with high MHz or HBM - Fiji needs HBM, GDDR5 wouldn't suffice I think, because even 1750 MHz 512 bit is way below 512 GB/s bandwidth, what Fiji now has with HBM and still scaling with it if you overclock it, what essentially proves that HBM is worth it, even just for the extra bandwidth it delivers, set aside the power reduce which made Fiji/Fury X possible at all.

It is a improvement, every bit helps. Nothing to do with "weaseling out", rofl, you simply don't understand my point, thats all. Again: HBM has high bandwidth, Fiji needs it, and overclocking HBM makes it even better - so what you said there on Page 1, about HBM being useless, is utter bullshit. And my overclocking-argument was just to prove how wrong your statement is, has nothing to do with scaling, what I know myself of, isn't the best, when overclocking HBM - but still it helps. And it is very reasonable, because you don't need more voltage for that +50-100 MHz on HBM, but you get free additional performance.

No your point is invalid. Just because it doesn't scale linearly doesn't mean its not bandwidth starved. AMD themself always speak of "how bandwidth limited cards are these days" and "always need more" - so AMD is wrong? Nope. Diminishing returns doesn't mean it is not bandwidth starved. Your point is just an unproven opinion of yours, nothing more.


Wrong.

Nope I'm very likely right with this. Without a possible Fury X, or say, a card that is strong enough to go face to face with the Titan X or 980 Ti, AMD would have scraped the entire line. Just "Nano" isn't worth it to do the Fiji+HBM+Interposer etc. thing, because Nano is not a very good selling card at all. Fury X is the first card - everything else is just a alternation to it, that's a 99,9999% fact, so if you delete Fury X because it's not possible without HBM, because TDP would be too high (this is a fact), you delete the entire Fury / Nano line. I thought that at this point everyone understood this - was written in articles everywhere, how important HBM was to design the Fiji/Fury-line of cards. Here on TPU, too, I think.

What? I just see a lot of senseless words here. Probably you didn't understand my point (again) and answered with some weird silly stuff. Plus, HBM is AMD tech, not Nvidia, you're completely off track here. HBM1 is a AMD exclusive, if AMD didn't want to make Fiji, HBM1 would not exist now - they would have skipped it for HBM2. But again - I don't think you understood my point. My point was, a Fury X with 275W TDP was only possible with HBM, not with GDDR5. Simple.


Yep, and I call your bluff or to say it in german words - you fell out of the window, you opened too wide and too fast. You're completely wrong on all or almost all points, certainly the most important ones about HBM and Fiji/Fury X.
Your psycho games are just utter childish bullshit, sorry. And I could laugh my ass off of you trying to insult me or call me a idiot. The idiot here is you, and it's not the first time I see you pissing off people or bullying around, or just being senselessly arrogant about just everything. And your other weakness is, that you just take about everything personal - THIS is silly. You're like that "Sony" guy, just somewhat alternated. Fact is, your pissing people off, because you're unable to debate in a civilized / friendly manner. Continue to do this and you're the first person on my black list, congratulations weirdo.


As I've understood it, GDDR5X is just a small evolution of GDDR5 and is easy to implement, saying, it does maybe not need new controllers etc. but I'm not sure about this, so take my answer with a grain of salt. I'm only sure HBM/HBM2 is expensive, because of the way it needs to be connected on the interposer and many companys working together to implement it.


GDDR5X is just a small evolution, not a revolution so I don't see the problem here. HBM/HBM2 is the big story - and will in 2016 still be. But AMD has that already behind them at least partially.


I never said "90%" of cards. I said only (semi-)highend cards will have HBM2. That's all. And I didn't say "they aren't starting on HBM2 chip orders", thats just a wrong assumption you made.


I don't care what you think of me. You're certainly the most arrogant person I've seen in my entire life and I'm on the internet since I was 13 and I'm 29 years old, so it means something you're winning that "prize". As I already stated, if you continue your misguided behaviour towards me, you're the first person on my blacklist and I won't miss you for one second. And your arguments are already proven as wrong or misguided, so there's no point in what you try to do here.


Well, I don't know what your parents did to you, but you seem to had a hard childhood. Life is about joy, not the hatred you are practicing. Well maybe not hatred but clearly in the direction of it. I would rather call it "disliking behaviour" or something of that sort. So you're an ass and you're happy with it and you try to make it legal by thinking its just the best way to be. So heres the surprise: it is NOT. Your behaviour is stressing everyone, a lot of your arrogance is based on you thinking you're extremely smart, but almost everything you said is extremely replacable or simply wrong. So continue your senseless ignorance - you are still wrong with your behaviour, and you can argue about that your entire life, you still will be wrong. Your behaviour is just bad. I tried to be tolerant towards you, when I first saw you quoted me and started the discussion with me, but I already knew you're difficult. But now you reached a point where my tolerance ends, I don't need random persons in the internet calling me a idiot or silly (or both), and then after it acting as it would be "justified". There are billion times more intelligent persons wandering the earth than you, and they are still humble and friendly. Think of that for a second if you CAN. You're just pitiful, sorry.

I'll be blunt, because it seems like there's a language barrier. You're changing the definition of words to suit your goals, and have yet to prove any of your points with anything more than inductive reasoning.

I'll take the direct insults to me as a sign that you have nothing to contribute.

Just for giggles though, I'm going to point out one more discrepancy that you seem to overlook. It seems to highlight exactly where your logic jumps the tracks.


Your original quote:

...
Well that's no problem. If it arrives with HBM2, they can plan for it and produce cards with it (GDDR5X). I don't see why this would be a problem. And as I said, I only see HBM2 on highend or highest end (650$) cards. I think the 400 (if any)/500/650$ cards will have it - so everything cheaper than that will be GDDR5X, and that's not only "cheap cards". I don't see 200-400 as "cheap". Really cheap is 150 or less and this is GDDR5-land (not even GDDR5X) I think.

You then claim:

..
I never said "90%" of cards. I said only (semi-)highend cards will have HBM2. That's all. And I didn't say "they aren't starting on HBM2 chip orders", thats just a wrong assumption you made.
....

Do you propose that the market for $650 and up cards is more than 10% of the market? If you're not claiming that, then you are demonstrably wrong. Perusing fact land, something which you seem incapable of doing in the slightest, we see that counting every card listed, the top end cards ($650 and higher) don't even make up 10% of user hardware. Here's the link: http://store.steampowered.com/hwsurvey/directx/

Instead of showing any facts that prove my arguments invalid, you change the goal posts, you start guessing about my childhood and calling names. Those are the tactics of a two year old. HBM is absolutely vital for Titan level performance is argued down to HBM overclocking improves performance (no numerical results were given by you, and the ones I provided contradicted your assertion). HBM is only going to be in high end cards is stated, then you say that these supposed high end cards make up more than 10% of the market. AMD absolutely needs HBM, yet the 380 cards are functionally just a minor upgrade of the 7970 and they compete reasonably with the 970 offering from Nvidia (I'm not linking results here, because it's honestly more effort than you've seen fit to do thus far).

You've made some bold assertions here, without doing any leg work to back them up. Whenever confronted with the fact that HBM wasn't necessary for Titan level performance, as demonstrated by the Titan running on GDDR5, you claim I don't understand how memory works. I'm willing to chalk this up as interpretation differences, but you run with the idea and negate everything else said with the argument that "he must not understand bandwidth, and therefore must not be able to comprehend HBM versus GDDR5."

I've openly admitted to being less than technically precise. Technically the HBM conversion saves about 15% of the die (25% die area with GDDR5, rather than 10% with HBM http://semiaccurate.com/2015/06/22/amd-talks-fiji-fiji-x-odd-bits-tech/). This means that because 29 mm^2 is not 15% of 625 mm^2 (596 actual less 625 for maximum size at fab) you would lose some performance. At the same time, you are unwilling to admit that Nvidia demonstrated HBM is not necessary for this level of performance. You aren't even willing to agree that AMD is setting artificial thermal envelopes for their cards. I don't get how you can be so stubborn in believing your point is unassailable without any facts, and with the facts others provide basically proving your point invalid.


I'd like to quote your exact post, so no editing by you can hide the meltdown and insults you've created. Here it is, in its full glory.

...
I don't care what you think of me. You're certainly the most arrogant person I've seen in my entire life and I'm on the internet since I was 13 and I'm 29 years old, so it means something you're winning that "prize". As I already stated, if you continue your misguided behaviour towards me, you're the first person on my blacklist and I won't miss you for one second. And your arguments are already proven as wrong or misguided, so there's no point in what you try to do here

Well, I don't know what your parents did to you, but you seem to had a hard childhood. Life is about joy, not the hatred you are practicing. Well maybe not hatred but clearly in the direction of it. I would rather call it "disliking behaviour" or something of that sort. So you're an ass and you're happy with it and you try to make it legal by thinking its just the best way to be. So heres the surprise: it is NOT. Your behaviour is stressing everyone, a lot of your arrogance is based on you thinking you're extremely smart, but almost everything you said is extremely replacable or simply wrong. So continue your senseless ignorance - you are still wrong with your behaviour, and you can argue about that your entire life, you still will be wrong. Your behaviour is just bad. I tried to be tolerant towards you, when I first saw you quoted me and started the discussion with me, but I already knew you're difficult. But now you reached a point where my tolerance ends, I don't need random persons in the internet calling me a idiot or silly (or both), and then after it acting as it would be "justified". There are billion times more intelligent persons wandering the earth than you, and they are still humble and friendly. Think of that for a second if you CAN. You're just pitiful, sorry.

My parents taught me that whenever I had a tantrum I lost the argument. They taught me to question my facts, and to question my own values. I'm going to bold the sections where you can't even be bothered to keep consistent standards when you're insulting someone instead of attacking their argument.

Let's also highlight your tolerance. You claim your actions are especially just, yet here's what started this discourse.

Can you all not read? There's written on the presentation that it has reduced voltage, so power usage still goes down, not up - the straight opposite....

If you start your conversation by assuming everyone else is an idiot, perhaps you've been the first to throw a stone.

Perhaps initially stating that the only way Fiji could have shown Titan level performance was with HBM, then changing the argument mid way through to HBM overclocking increasing performance without any figures (never even highlighting any numbers, but I'm going to be nice and provide you 512 GB/s for Fury http://www.forbes.com/sites/jasonev...-new-fiji-graphics-card-beats-nvidias-980-ti/ and 336.5 for Titan X http://www.geforce.com/hardware/desktop-gpus/geforce-gtx-titan-x/specifications), and then further changing the argument to HBM not being implemented on anything but the most high end cards was not a valid tactic to make an argument with. Colloquially, that's referred to as nailing gelatin to a wall. When there is no stationary target you can't make an argument for it. It's a great tactic for someone who doesn't want facts to get in the way of narrative. It must have been HBM, because I say so.

Perhaps your lack of a grasp on English is genuinely an issue (I know my grasp on German has often made it difficult to make a salient point consistently).

Maybe, just maybe, you lack the ability to have a discussion. The internet is great for people who want safe spaces, where their opinion is gold. If you'd like to make assumptions about me, I believe the favor should be returned though. After a year in Lubeck, I realized that most Germans loved saying they had free speech, until something damning was said. The US may have throw-backs like the Westboro baptists and KKK, but we don't stop peaceful demonstrations because their ideology is unappealing (look up the socialist party in Lubeck, they do an annual march which during my time was completely non-violent but the police acted like they were laying siege to the town). I can only speak to the Germans on this issue, but after a year it was consistent for most people (80% in the 20-30 person sample sizes I could accurately gauge, though there definitely were counter examples).

I'd actually implore you to block me. It's one less person who can claim to be a victim whenever they can't be bothered to support their arguments. One less person who cannot admit if they've made a mistake, and have to be right so they change the arguments to fit their narrative.

After saying all of this though, it is funny to see your response. Rather than stand on a good argument with facts (I see no citations for any of your points), you choose to insult me personally. I applaud the meltdown, because it's truly worthy of a smile. Also, I'd like my prize to be a cookie. Where's my cookie?

 

[Kanan]

I read and answer directly (because I don't want to read it 2 times), so don't wonder if my opinion about you changes in the post, progressing further.

I'll be blunt, because it seems like there's a language barrier. You're changing the definition of words to suit your goals, and have yet to prove any of your points with anything more than inductive reasoning.

Language barrier? Always a good excuse if you got no arguments to counter someone. But use everything you have, you probably need it.

I'll take the direct insults to me as a sign that you have nothing to contribute.

Nope, that only shows to me that your logical thinking is inferior. That's weak. The person with the insults was you, not me. I just got to a point that I couldn't stand your foolish attacks on me anymore and started to counter it with "fire against fire". I have more to contribute than you think. But underestimate me, I like that, because that way I can only win.

Do you propose that the market for $650 and up cards is more than 10% of the market? If you're not claiming that, then you are demonstrably wrong. Perusing fact land, something which you seem incapable of doing in the slightest, we see that counting every card listed, the top end cards ($650 and higher) don't even make up 10% of user hardware. Here's the link: http://store.steampowered.com/hwsurvey/directx/

Everybody knows by now that your childish attacks on me ruined the discussion, this is why a big part of my post will just be an answer to that. But I will answer it nonetheless:
What I meant in my first post about it was "Highend and Highest End cards", thats 300/400/500/550/650/1000$ cards (for me) - that's what I meant. Just to clarify that. If that is just 10% of the market, then yes, you are right. But it's more than 10%, afaik.

Instead of showing any facts that prove my arguments invalid, you change the goal posts, you start guessing about my childhood and calling names. Those are the tactics of a two year old.

You aren't showing facts either, no one here does. This topic is highly speculative and that's more than obvious - we talk about future tech here. But tell me, for what you want proof other then the things I can't know about and I'll deliver it. Fury X, Fiji, HBM1 etc. - what I said about these can be proven, there is info on that in the internet, I read it myself, I'm not making up things, I assure you that.
And btw. "calling names"? Is that meaning insulting someone? That's YOU who started that (and by far not the first time). Seems you don't know what "selfreflection" means, and I did that for you, because you're unable to do it. Not my fault that you are acting like a child, though. Is the person who is always insulting other ones because he is aggressive and feels "attacked" by just informations or neutral answers, me or you, here? I just answered on your insults, after you did that a few times against me - the first times I just ignored it. Everyone has a limit in what he or she can stand. Really, you're like a child - harassing me like 100 times and then after I react to it, reacting like that, here, acting as the one who is "innocent". Sorry, thats just pitiable what you do here.

HBM is absolutely vital for Titan level performance is argued down to HBM overclocking improves performance (no numerical results were given by you, and the ones I provided contradicted your assertion).

Can you imagine a Fury X without HBM and with GDDR5, but still a TDP of only 275 W who can cope with the Titan X? No. And overclocking is another topic, you are mixing things here. I never said overclocking the HBM helps much on the Fiji cards, though. I just said "it does" to prove that they are not bandwidth saturated. And again: AMD themselves said that. Perhaps they are "silly idiots" too, like you said?

AMD absolutely needs HBM, yet the 380 cards are functionally just a minor upgrade of the 7970 and they compete reasonably with the 970 offering from Nvidia (I'm not linking results here, because it's honestly more effort than you've seen fit to do thus far).

The 380 has no chance against 970. That's a misinformation. See here: http://www.pureoverclock.com/Review-detail/msi-r9-380-gaming-4g-review/7/
You probably meant the 960. And why AMD needs HBM1 me and HumanSmoke already told you X times - are you that ignorant? Again: they needed it, so the Fury X was possible with a core clock of 1050 MHz at TDP of 275W. You can read that in reviews world wide, or do you need a link for that too?

You've made some bold assertions here, without doing any leg work to back them up. Whenever confronted with the fact that HBM wasn't necessary for Titan level performance, as demonstrated by the Titan running on GDDR5, you claim I don't understand how memory works.

Well, we try it your way: it is not, yes, but it would be impractical. A Fury X (1050 MHz clock, 4096 Shaders) would suck way over 300W, with GDDR5 of 1500 MHz or more on a 512 bit bus. You accept that, or do you not? For me that is a clear fact, and informations on that, again, you can find world wide in reviews. That was my point the whole time, perhaps you understood me wrong, or I understood you wrong, I don't know - talking to you is not easy, at least for me.

I'm willing to chalk this up as interpretation differences, but you run with the idea and negate everything else said with the argument that "he must not understand bandwidth, and therefore must not be able to comprehend HBM versus GDDR5."

I'm trying to be reasonable with you and tolerant, or else I wouldn't answer here.

I've openly admitted to being less than technically precise. Technically the HBM conversion saves about 15% of the die (25% die area with GDDR5, rather than 10% with HBM http://semiaccurate.com/2015/06/22/amd-talks-fiji-fiji-x-odd-bits-tech/). This means that because 29 mm^2 is not 15% of 625 mm^2 (596 actual less 625 for maximum size at fab) you would lose some performance.

Yep I concur with you on that one, but don't forget that HBM hindered Fiji going all out on ROPs, because the chip size was limited by it (HBM next to the chip, limited space on the Interposer - and the Interposer generally). That's the problem. Fiji only has 64 ROPs and I'm sure they wanted to do more, but haven't got the space to do it. Fiji with 96 ROPs would have been so nice, probably even faster than custom 980 Tis and a lot faster than reference ones. A balanced chip it would have been - sadly it is not.

At the same time, you are unwilling to admit that Nvidia demonstrated HBM is not necessary for this level of performance.

Where did I say that? I just spoke about AMD, I didn't mean Nvidia with that. Clearly Maxwell is better than GCN 1.2 and the 980 Ti the best card on the planet. Not only because it is efficient - because it overclocks very good too, and leaves everything in the dust, nearly coming to performance of a 295X2, with half the power needed to achieve it.

You aren't even willing to agree that AMD is setting artificial thermal envelopes for their cards. I don't get how you can be so stubborn in believing your point is unassailable without any facts, and with the facts others provide basically proving your point invalid.

Artificial thermal envelopes? I didn't really talk about that. Afaik I just said that, without HBM a Fury X as we have it now, with 1050 MHz core clock would have been impossible, that's all. Surely, it is artificial, but it's a very rational TDP too. More than 275W is just too much - AMD can't do that without emberassing themselves to Nvidia and the market. I mean, AMD (highend) cards as they are now (all besides the Nano, I'd say) aren't really efficient, but a TDP more than 275W would have just been an disaster, I'd say. They can't do that. You need some ecology nowadays, aside from it just beeing inefficient in a computer setup and on presentations for shareholders. The Fury X, as it is now, is still not efficient enough - but with GDDR5 ... wow, that would have been terrible.

I'd like to quote your exact post, so no editing by you can hide the meltdown and insults you've created. Here it is, in its full glory.

Nobody is perfect, and you are no angel either. And I don't have a problem with that, as you can see.

My parents taught me that whenever I had a tantrum I lost the argument. They taught me to question my facts, and to question my own values. I'm going to bold the sections where you can't even be bothered to keep consistent standards when you're insulting someone instead of attacking their argument.

Your parents told you that? For me this is common knowledge, the least everyone should do, nothing special. I started insulting you, after I had enough of your many insults against me - again. You seem to ignore that fact.

Let's also highlight your tolerance. You claim your actions are especially just, yet here's what started this discourse.

I never said I'm totally "just" or an angel. But, again, do what you want.

If you start your conversation by assuming everyone else is an idiot, perhaps you've been the first to throw a stone.

I never did that, and I'd never do that. You just read what you want into what I write. Just because I ASKED if "everyone can't read", I didn't call ANYONE an "idiot", you're totally exaggerating here. As I see it, it was a justified question, because a lot of people seemed to ignore the fact that GDDR5X is reducing voltage and therefore power needed. Perhaps I'm wrong on that point, but I wanted to point it out, nonetheless. That's all what it was about, not calling anyone an idiot. And you are the only one reacting extreme to it, as a matter of fact. I don't have intentions to call anyone a idiot or attack anyone - I just do that when someone else tries to do it with me. Maybe you read such bad things into what I write because you have that negative thinking yourself. People tend to do that. Just generally speaking.

Perhaps initially stating that the only way Fiji could have shown Titan level performance was with HBM, then changing the argument mid way through to HBM overclocking increasing performance without any figures (never even highlighting any numbers, but I'm going to be nice and provide you 512 GB/s for Fury http://www.forbes.com/sites/jasonev...-new-fiji-graphics-card-beats-nvidias-980-ti/ and 336.5 for Titan X http://www.geforce.com/hardware/desktop-gpus/geforce-gtx-titan-x/specifications), and then further changing the argument to HBM not being implemented on anything but the most high end cards was not a valid tactic to make an argument with. Colloquially, that's referred to as nailing gelatin to a wall. When there is no stationary target you can't make an argument for it. It's a great tactic for someone who doesn't want facts to get in the way of narrative. It must have been HBM, because I say so.

I tried to clarifiy that, now. Maybe I'm still too dumb for english. Or maybe you and your understanding is incompatible with me, because I don't seem to have problems with others so far in this thread. Fact is I didn't want to change my arguments, or the topics, or what I was trying to talk about. I just tried to drive a discussion and defend my points, that's all.

Perhaps your lack of a grasp on English is genuinely an issue (I know my grasp on German has often made it difficult to make a salient point consistently).

Maybe. My english is far from perfect and in discussions with advanced english needed it gets somewhat hard for me to say what I want to say, in english. In german I would do that in sleep...

Maybe, just maybe, you lack the ability to have a discussion. The internet is great for people who want safe spaces, where their opinion is gold. If you'd like to make assumptions about me, I believe the favor should be returned though.

I thought exactly that about you too, so in my view, you wouldn't be the exact person to criticize others about "discussing in a proper way".

After a year in Lubeck, I realized that most Germans loved saying they had free speech, until something damning was said. The US may have throw-backs like the Westboro baptists and KKK, but we don't stop peaceful demonstrations because their ideology is unappealing (look up the socialist party in Lubeck, they do an annual march which during my time was completely non-violent but the police acted like they were laying siege to the town). I can only speak to the Germans on this issue, but after a year it was consistent for most people (80% in the 20-30 person sample sizes I could accurately gauge, though there definitely were counter examples).

First, idiots are everywhere. Second, don't try to generalize, you can only fail doing that - you are not God, you don't know enough to do that. Third, I don't see myself as a german, more like a human with some german influences, or rather international influences, most of it being american/german/european with some asiatic in it. So I don't have any problems when you criticize germans, as I see myself as a human, not a german per sé.
My experience is this: German mentality isn't easy at times, germans have problems in discussions, because they tend to be arrogant at times, or have a eager will to be right. I had that problem too when I was younger. Germans tend to generalize a lot too. But what is good about germans, it's part of their general mentality to be very precise. And I like that. This is basically why germans are such good engineers and nowadays humans too - but this is way too offtopic, just some opinions from me, an FYI if you want to know my view on it. What I like about americans (and this is why I'm here, not on a german forum), they are easier on things and less serious. And when they are speaking about things, they overreact rarely, basically trying to stay cool all the times. This is a fat + when debating. Just my opinion though, nothing more or less.

I'd actually implore you to block me. It's one less person who can claim to be a victim whenever they can't be bothered to support their arguments. One less person who cannot admit if they've made a mistake, and have to be right so they change the arguments to fit their narrative.

I won't block you just yet, maybe we can come to an agreement after all, you seem to be reasonable at times at least, and you seem to try to understand me better.

After saying all of this though, it is funny to see your response. Rather than stand on a good argument with facts (I see no citations for any of your points), you choose to insult me personally. I applaud the meltdown, because it's truly worthy of a smile. Also, I'd like my prize to be a cookie. Where's my cookie?

You're a funny guy, I give you that. Again, I tried to clarify things for you, I hope it is better now for both. haha, imagine I gave you a cookie, it's in your hand just open it.

 

[Aquinus]

Wall of text

I suggest that you stop acting like a child and let it go (cue Frozen.) Given your entire post you've introduce very little new information and are simply trying to defend yourself by repeating the same thing you said with added levels of frustration at every turn. If you two really have a problem, settle it in PMs but, at this point, all I'm hearing is @lilhasselhoffer getting irritated (yes, he most definitely know his shit, like @HumanSmoke as well) but, all of your posts have been generalizations without any real facts with some links interspersed through your posts (rarely). @lilhasselhoffer has offered about 4 to 5 times more data than you have, so I suggest getting pissed off a little less and re-reading his posts again. I'm not going to lie, most of the time when I read @lilhasselhoffer or @HumanSmoke posts, I re-read it to ensure I didn't get it wrong. These two guys tend to give huge amounts of information in one sitting.

@lilhasselhoffer knows when to admit that he was wrong about something and he did earlier in this thread, I think you (@Kanan ,) should take a step back for a little bit and take a breather to realize that most of this "anger" is in response to your attitude and how some of the things you're saying appear inconsistent.

Either way, this argument has evolved beyond the scope of this thread and I would like it to stop.

 

[lilhasselhoffer]

I suggest that you stop acting like a child and let it go (cue Frozen.) Given your entire post you've introduce very little new information and are simply trying to defend yourself by repeating the same thing you said with added levels of frustration at every turn. If you two really have a problem, settle it in PMs but, at this point, all I'm hearing is @lilhasselhoffer getting irritated (yes, he most definitely know his shit, like @HumanSmoke as well) but, all of your posts have been generalizations without any real facts with some links interspersed through your posts (rarely). @lilhasselhoffer has offered about 4 to 5 times more data than you have, so I suggest getting pissed off a little less and re-reading his posts again. I'm not going to lie, most of the time when I read @lilhasselhoffer or @HumanSmoke posts, I re-read it to ensure I didn't get it wrong. These two guys tend to give huge amounts of information in one sitting.

@lilhasselhoffer knows when to admit that he was wrong about something and he did earlier in this thread, I think you (@Kanan ,) should take a step back for a little bit and take a breather to realize that most of this "anger" is in response to your attitude and how some of the things you're saying appear inconsistent.

Either way, this argument has evolved beyond the scope of this thread and I would like it to stop.

Agreed. I won't be responding again.


Edit:
Let me be clear, I have sent a PM. I will no longer be responding to Kanan period.

 

[Kanan]

I suggest that you stop acting like a child and let it go (cue Frozen.) Given your entire post you've introduce very little new information and are simply trying to defend yourself by repeating the same thing you said with added levels of frustration at every turn. If you two really have a problem, settle it in PMs but, at this point, all I'm hearing is @lilhasselhoffer getting irritated (yes, he most definitely know his shit, like @HumanSmoke as well) but, all of your posts have been generalizations without any real facts with some links interspersed through your posts (rarely). @lilhasselhoffer has offered about 4 to 5 times more data than you have, so I suggest getting pissed off a little less and re-reading his posts again. I'm not going to lie, most of the time when I read @lilhasselhoffer or @HumanSmoke posts, I re-read it to ensure I didn't get it wrong. These two guys tend to give huge amounts of information in one sitting.

@lilhasselhoffer knows when to admit that he was wrong about something and he did earlier in this thread, I think you (@Kanan ,) should take a step back for a little bit and take a breather to realize that most of this "anger" is in response to your attitude and how some of the things you're saying appear inconsistent.

Either way, this argument has evolved beyond the scope of this thread and I would like it to stop.

Nope, this is the 2nd time you are going straight against me, never really understanding my point. Me childish? Then you are ignorant, far worse I'd say. And still, I'am not who acted childish here in the first place, only my reaction to his Kindergarten was somewhat childish, I give you that - but I also already knew that, I don't need your critique on my person, I know what I do pretty much better than you ever will. Plus, you're pretty biased I'd say, too, because you know him way longer than me. This is more than obvious. Either way you interfered in the discussion without any real sense, I think it was on the way to get better, before you interfered here. You act wise and grown up, but I don't accept your points. A lot of blabla and some informations I already knew (every human being is childish at times, you are no exception to that, and if yes, you are limited I'd say - being childish is good sometimes, and everyone should have that in his or her life).

Next thing is, anger is never justified. Yes I was angered too by his accusations and his childish bullshiting, and this was my mistake too. But fact is, being angered and then go to someone and harass him that way he did, is NEVER the right way - you didn't realize that, which further proves my point that you are very well biased to him and / or not wise and grown up enough to make a point here. Move along, I don't accept your authority. This case is closed, now - and you made it worse than it was before.

Agreed. I won't be responding again.

For gods sake, thank you.

 

[lilhasselhoffer]

Alow me a little thought exercise.

1) Is HBM absolutely necessary on high end cards?
Demonstrably, it is not. Right now the implementation of GDDR5 and HBM have similar levels of performance.

2) Will HBM be necessary in the future?
Absolutely. HBM saves about 10% of the GPU die space, by having the controller be much smaller. While this does introduce the need for an interposer, it allows you to use more of the die area for the GPU. If you aren't really redesigning your architecture, then you've got to have more GPU area to overcome the lack of increasing performance.

3) Is HBM a necessity on Fiji?
Yes and no. There art limitations from the GPU fabricator, that limit die size. AMD, not investing resources in refining their GPUs, has placed their bets on a new type of memory to allow for more of the die to be used by the GPU. Because of this, AMD requires HBM to get Titan level performance. The counter argument is Titan itself. With the same die size restrictions, and a GDDR5 memory controller, Titan proves that it isn't necessary to have HBM in order to get the performance of Fury. To be honest though, we've started with a lot of assumptions. Fury set performance targets based upon Titan. It set thermal limits based upon cooling. It even determined size artificially (there's anything from half height to triple height cards out there). AMD could have altered any one of these goals and made a very different Fiji.

4) Is HBM really a stumbling point for Fiji?
Demonstrably, no. There is not a 1:1 correlation between overclocks on HBM and raw performance of the card. This is demonstrable via overclocking figures, and thus proves the additional bandwidth provided to Fiji via HBM isn't as significant a limiter as the GPU itself (nearly 1:1 ratio on GPU overclocking performance)

5) Will GDDR5X influence anything (Arctic Islands versus Pascal)?
Not really. Nvidia and AMD both likely already have signed contracts that they've negotiated for parts. You could theoretically argue that they could redesign the memory controller, which would be a huge time and resource investment this late in the game. If the engineering hurdles were overcome quickly, they'd still have to contend with existing stocks of GDDR5 (not to mention proving out suppliers for the new memory standard). You could use regular GDDR5 with a GDDR5X controller (part of the supposed backwards compatibility), but then why spend the money? You can;t just switch over to better product later, customers would ditch your cards when they found out that they paid the same price for a demonstrably worse product. It's more likely that GDDR5X will be present in a line refresh, once DDR5 stocks disappear. This puts it well into 2017. This time window agrees with proving out a completely new product, while not impacting the release of a dofferent and experimental new technology (HBM2)

A whole conversation without conjectures about the author, or using uncommon or fluid definition of words. Sigh.

 

[cadaveca]

Alow me a little thought exercise.

No. Here's why:

HBM is die stacking. The interposer does not replace a memory controller. It replaces traces in PCB that connect the memory controller to the memory. It did not affect die space. It affected PCB size (see the small fury PCBs). HBM allows this because multiple memory chips are stacked on top of each other. The tech that makes this possible is not the interposer... that interposer is really not important to anything other than that it allows the TSVs of HBM to work right.

Having less interconnects to connect the memory does affect die space a small bit, but not in a very large way. It actually made things far more complex than they needed to be. But, AMD's success with this design grants wins in other markets. Not discrete GPUs.

 

[Aquinus]

Nope, this is the 2nd time you are going straight against me, never really understanding my point. Me childish? Then you are ignorant, far worse I'd say. And still, I'am not who acted childish here in the first place, only my reaction to his Kindergarten was somewhat childish, I give you that - but I also already knew that, I don't need your critique on my person, I know what I do pretty much better than you ever will. Plus, you're pretty biased I'd say, too, because you know him way longer than me. This is more than obvious. Either way you interfered in the discussion without any real sense, I think it was on the way to get better, before you interfered here. You act wise and grown up, but I don't accept your points. A lot of blabla and some informations I already knew (every human being is childish at times, you are no exception to that, and if yes, you are limited I'd say - being childish is good sometimes, and everyone should have that in his or her life).

Next thing is, anger is never justified. Yes I was angered too by his accusations and his childish bullshiting, and this was my mistake too. But fact is, being angered and then go to someone and harass him that way he did, is NEVER the right way - you didn't realize that, which further proves my point that you are very well biased to him and / or not wise and grown up enough to make a point here. Move along, I don't accept your authority. This case is closed, now - and you made it worse than it was before.


For gods sake, thank you.

I asked to you take take it to a PM as this isn't the appropriate setting and like other members, I'm sick of hearing your whining and back peddling so I simply asked nicely for you to stop. Yet you come around and attack me and call me childish because I want it to end? Thanks @Kanan, you just confirmed that you're not listening to anyone. Maybe we need a moderator to step in because you're clearly not going to listen to anyone other than yourself because you simply redirect your anger based on whose replying to you.

My recommendation still stands, bring this to a PM and drop it. I'm sure a mod would agree with me given the attitudes.

HBM is die stacking. The interposer does not replace a memory controller. It replaces traces in PCB that connect the memory controller to the memory.

I was under the impression that HBM's memory controller is indeed smaller than its GDDR5 counterpart. Size of the IMC has nothing to do with the interposer or where the contacts for the memory controller are but rather the control logic for utilizing the DRAM. Maybe HBM is more simple by design because it's not clocked as high so things like clock buffers might be less important. That's all a guess though, I haven't looked too deeply into it.

 

[RCoon]

I'd say a good 70% of the entire content of this thread is total trash - incessant bickering and not in any way related to the subject at hand. Take it to PMs, or just use the ignore button. That's what it's there for.

First and last warning.

 

[HumanSmoke]

Having less interconnects to connect the memory does affect die space a small bit, but not in a very large way.

It isn't just the memory controller that is simpler and smaller, it is also the memory physical layer. I wouldn't call 15% smaller "a small bit" - especially on a mid range GPU where the memory interface takes up a larger proportion of the whole die area - and lets face it, this is the area where HBM will be the hard sell. Even using the only example so far, Fiji, the difference between HBM and GDDR5 is the difference (taking @lilhasselhoffer 's SA link as gospel) between manufacturable (596mm²) and borderline non-manufacturable (646mm²) - unless you fancy AMD's chances of having Fiji fabbed by Intel, because 646mm² is certainly uncomfortable for TSMC outside of simple chips (FPGA, CMOS and CCD sensors and the like).

HBM allows this because multiple memory chips are stacked on top of each other.

There is actually a pretty good chance that GDDR5/5X will also follow suit. Conventional DDR modules are already there.

 

[cadaveca]

It isn't just the memory controller that is simpler and smaller, it is also the memory physical layer. I wouldn't call 15% smaller "a small bit" - especially on a mid range GPU where the memory interface takes up a larger proportion of the whole die area - and lets face it, this is the area where HBM will be the hard sell. Even using the only example so far, Fiji, the difference between HBM and GDDR5 is the difference (taking @lilhasselhoffer 's SA link as gospel) between manufacturable (596mm²) and borderline non-manufacturable (646mm²) - unless you fancy AMD's chances of having Fiji fabbed by Intel, because 646mm² is certainly uncomfortable for TSMC outside of simple chips (FPGA, CMOS and CCD sensors and the like).

There is actually a pretty good chance that GDDR5/5X will also follow suit. Conventional DDR modules are already there.

15% smaller for the memory controller only, simply due to less connections to the memory ICs. 15% of 25% is a grossly small number. And yeah, HBM doesn't have much place on mid-range or lesser at the moment, but it does allow for pretty compact GPUs that have a great position in the console/BGA market, and in the future, it WILL make its way into smaller GPUs, with less HBM memory chips (4GB in a single die?), and smaller interposers.

I was under the impression that HBM's memory controller is indeed smaller than its GDDR5 counterpart. Size of the IMC has nothing to do with the interposer or where the contacts for the memory controller are but rather the control logic for utilizing the DRAM. Maybe HBM is more simple by design because it's not clocked as high so things like clock buffers might be less important. That's all a guess though, I haven't looked too deeply into it.

As I said, if the memory controller is 25% of the GPU, 15% of that is like what... 3.75% of the entire GPU space saved? That's pretty small. Even at 50%, 15% save of 50% is only 7.5% of the entire GPU. It simply allows for less interconnects to the memory ICs, and this is how the space is saved. Think about how many memory ICs are on 290X/390X. Remove all of those chips but 4, and you can easily imagine how less complex the PCB itself is, never mind how many fewer interconnects are needed.

 

[lilhasselhoffer]

No. Here's why:

HBM is die stacking. The interposer does not replace a memory controller. It replaces traces in PCB that connect the memory controller to the memory. It did not affect die space. It affected PCB size (see the small fury PCBs). HBM allows this because multiple memory chips are stacked on top of each other. The tech that makes this possible is not the interposer... that interposer is really not important to anything other than that it allows the TSVs of HBM to work right.

Having less interconnects to connect the memory does affect die space a small bit, but not in a very large way. It actually made things far more complex than they needed to be. But, AMD's success with this design grants wins in other markets. Not discrete GPUs.

Maybe I didn't make the point clear. The memory controller for HBM is largely moved off of the GPU die, into the interposer. As per the estimation provided here (http://semiaccurate.com/2015/06/22/amd-talks-fiji-fiji-x-odd-bits-tech/), that means an additional 15% of the GPU die doesn't have to be allocated for the memory controller (same link cites 25% of the Hawaii die is GDDR5 controller). Your interposer will have to be a complex web of interconnects, but they've apparently already decided to make up for this by using manufacturing lithography that is a couple of generations old. You're therefore capable of using much more of the GPU die for the GPU.


This says nothing about interposer limitations. I honestly look at this as "solving" the manufacturing issue by introducing a complicated web of connections between the GPU and its memory. You'll note that I've already said Nvidia has demonstrated that Titan level performance doesn't require such a complicated solution. Nvidia went back to the drawing board, and optimized their GPU design. AMD said screw optimization, let's just free up die space by making everything immensely more complicated. Only time will tell what solution is actually best, though it honestly doesn't matter until we've got something beyond a $650 solution.

Edit:

15% smaller for the memory controller only, simply due to less connections to the memory ICs. 15% of 25% is a grossly small number. And yeah, HBM doesn't have much place on mid-range or lesser at the moment, but it does allow for pretty compact GPUs that have a great position in the console/BGA market, and in the future, it WILL make its way into smaller GPUs, with less HBM memory chips (4GB in a single die?), and smaller interposers.



As I said, if the memory controller is 25% of the GPU, 15% of that is like what... 3.75% of the entire GPU space saved? That's pretty small. Even at 50%, 15% save of 50% is only 7.5% of the entire GPU. It simply allows for less interconnects to the memory ICs, and this is how the space is saved. Think about how many memory ICs are on 290X/390X. Remove all of those chips but 4, and you can easily imagine how less complex the PCB itself is, never mind how many fewer interconnects are needed.

Please read the link again. 25% of the Hawaii die is controller. 10% of the Fiji die is memory controller. 15% is the savings of the raw die space, not 15% of the 25%.

Additionally, could you cite some sort of point? It's difficult to argue when somebody says "no" without framing what they are saying no to.

 

[cadaveca]

IN none of the articles you have linked does it say the memory controller itself is on the interposer die. That is the root of my "no". Maybe I missed it, but I'm pretty sure I didn't. Perhaps you can point that out to me?

From your recent link:

With HBM in Fiji, it is a lot simpler even though it is wider and has a net higher bandwidth. Now only about 10% of the die is the HBM controller, roughly 60mm^2, even though bandwidth goes up significantly saving both power and obviously expensive silicon.

So, the memory controller is nearly 50% smaller (110mm to 60mm). That's a different way of relating numbers than you did. Anyway, again, the interposer is NOT the memory controller.

That's what I'm looking for...the information that the interposer contains the memory controller. My point is that the controller is smaller because of less interconnections to memory ICs (4 chips HBM vs 16 chips of GDDR5).

The rest of what you have posted about... I don't care about. I don't care about NVidia vs AMD comparisons, or conjecture about what they may do, I care about what they have already done.

 

[HumanSmoke]

As I said, if the memory controller is 25% of the GPU, 15% of that is like what... 3.75% of the entire GPU space saved?

No it isn't.
The article, thoughtfully provided once again by @lilhasselhoffer , states that the 8 dual channel IMC's of Hawaii take up around 110mm² while the 8 dual channel IMC's of Fiji account for approximately 50mm²

IN none of the articles you have linked does it say the memory controller itself is on the die. That is the root of my "no". Maybe I missed it, but I'm pretty sure I didn't. Perhaps you can point that out to me?

Which memory controllers are you talking about?
The Hawaii IMC's...........................................................................................................................................................or Fiji's IMC's ???

.................

 

[cadaveca]

LuLz. You didn't wait for my edits. I have yet to see anywhere that the IMC is on the interposer. It's mentioned that it is passive (ie, not powered), so that idea that memory control is on the interposer doesn't make sense to me personally. It's also cheap, since it's passive, and just a few layers thick. It IS large, though, so could be fairly complex, but that's not how I read what is posted.

Also, HBM2 apparently removes the need for the interposer. PASCAL cards shown already have a rather normal-looking GPU substrate. So any conjecture about the interposer on future products that are reported to use HBM2 is useless. This also plays into the comparison of HBM vs GDDR5X.

Please read the link again. 25% of the Hawaii die is controller. 10% of the Fiji die is memory controller. 15% is the savings of the raw die space, not 15% of the 25%.

But above, you posted that the savings in die space is 10%? So I posted a bunch of other meaningless numbers like you did?

Is it 15%, or 10%????

2) Will HBM be necessary in the future?
Absolutely. HBM saves about 10% of the GPU die space, by having the controller be much smaller.

EDIT:

So, more to my point;

Macri explained that the interposer is completely passive; it has no active transistors because it serves only as an electrical interconnect path between the primary logic chip and the DRAM stacks.

http://techreport.com/review/28294/amd-high-bandwidth-memory-explained


No IMC on interposer.