AMD Pulls Radeon "Vega" Launch to October

[Kanan]

There is some truth to this rumor. Vega is rumored to be around 380mm^2. AMD's approach to high performance is in the case of Vega is 60% more shaders than the fully-enabled Polaris chip. Vega will definitely be able to at least compete with the GTX1080, which is what AMD need sto do real soon to start selling high profit GFX cards.


What Nvidia did with the GP104 is just unprincipled. If the performance rumors are true, then TitanX performance for $380 when a 390/GTX970 can play all your games at decent frame rates on max settings is is just unnecessary. Nvidia is trying to strangle AMD and squeeze them out of the high performance high profit GFX market.

Maybe. But I'm still thinking about Vega as a big chip, like GP100, because it uses HBM2 (720+ GB/s). A chip with 380mm² doesn't need expensive to produce (lot less $$ earning for AMD) HBM2, it could come with 256 or 384 or even 512 bit GDDR5X if needed. That's why I think this "news" is a hoax. Also why call it "Vega10"? "10" is the bigger chip, not the smaller. The Vega10 would have 4096 Shaders, whereas the Vega11 would have 3584 (like Fury X and Fury). This whole news makes no sense! Vega is 1 chip, meaning the Vega10 couldn't have 4096 shaders whereas the other has 6144, that's senseless bullshit. That 6144 core chip would obviously be another one. But you don't have to believe me, let's wait and see. hahaha ... "and so the stupid rumour's started again."

 

[G33k2Fr34k]

Maybe. But I'm still thinking about Vega as a big chip, like GP100, because it uses HBM2 (720+ GB/s). A chip with 380mm² doesn't need expensive to produce (lot less $$ earning for AMD) HBM2, it could come with 256 or 384 or even 512 bit GDDR5X if needed. That's why I think this "news" is a hoax. Also why call it "Vega10"? "10" is the bigger chip, not the smaller. The Vega10 would have 4096 Shaders, whereas the Vega11 would have 3584 (like Fury X and Fury). This whole news makes no sense! Vega is 1 chip, meaning the Vega10 couldn't have 4096 shaders whereas the other has 6144, that's senseless bullshit. That 6144 core chip would obviously be another one. But you don't have to believe me, let's wait and see. hahaha ... "and so the stupid rumour's started again."

Not likely. The Vega chip I'm talking about is probably going to use 2048 bit interface with four stacked memory chips, achieving 300 ish to 400 ish GB/s of memory bandwidth. This chip will likely be in the 490/490X cards. AMD won't built a 600mm^2 any time soon. It's just not practical.

 

[Kanan]

Not likely. The Vega chip I'm talking about is probably going to use 2048 bit interface with four stacked memory chips, achieving 300 ish to 400 ish GB/s of memory bandwidth. This chip will likely be in the 490/490X cards. AMD won't built a 600mm^2 any time soon. It's just not practical.

That's just a rumour you keep repeating I believe it after some substantial informations, or when it is fact indeed. Vega IS supposed to be that big chip btw. - or semi big at least. I don't see why they should do a smaller version Vega with HBM2 @ 2048 bit, instead of GDDR5X, it's just too expensive.

 

[G33k2Fr34k]

That's just a rumour you keep repeating I believe it after some substantial informations, or when it is fact indeed. Vega IS supposed to be that big chip btw. - or semi big at least. I don't see why they should do a smaller version Vega with HBM2 @ 2048 bit, instead of GDDR5X, it's just too expensive.

Yeah... I'm not sure about this whole Vega thing. GDDR5X is the more feasible memory solution, but it could be the case that using HBM would result in significant power savings, something AMD seem very concerned about with this generation. Also, we don't know how expensive HBM is in comparison to GDDR5X.

I believe we already know about an AMD graphics chip with 4096 SPs. And if the rumors are true, then it's a ~350mm^2 - 360mm^2 chip with HBM2 memory.

 

[Kanan]

Yeah... I'm not sure about this whole Vega thing. GDDR5X is the more feasible memory solution, but it could be the case that using HBM would result in significant power savings, something AMD seem very concerned about with this generation. Also, we don't know how expensive HBM is in comparison to GDDR5X.

I believe we already know about an AMD graphics chip with 4096 SPs. And if the rumors are true, then it's a ~350mm^2 - 360mm^2 chip with HBM2 memory.

I think GDDR5X costs A LOT less than HBM2, not even comparable. Just google how the Fury X is build (Interposer + HBM + GPU stuff), whereas GDDR5X are just better GDDR5 chips that are relatively cheap again. I won't comment on the other stuff again, we will see. Enough rumour fishing.

 

[Eroticus]

I think GDDR5X costs A LOT less than HBM2, not even comparable. Just google how the Fury X is build (Interposer + HBM + GPU stuff), whereas GDDR5X are just better GDDR5 chips that are relatively cheap again. I won't comment on the other stuff again, we will see. Enough rumour fishing.

Yes but, HBM2 is allooooooooooooooooooooooooooooooooooooooooooot faster than GDRR5X.

HBM1 -Fury X Memory bandwidth 512 (GB/s)
HBM2 - 1024 (GB/s) ?
GDDRX - 1080 Memory bandwidth 320 (GB/s)

 

[medi01]

Man, dafuq I was thinking:
http://www.techpowerup.com/221720/amd-to-launch-radeon-r7-470-and-r9-480-at-computex

So, only 470 and 480 they launch.

Say 470 is at 380x levels, 480 at 390x levels.
Both slower than 1070

Sure, that Vega chip must come earlier.

PS

During a recent interview, Roy Taylor – AMD / RTG’s head executive for Alliances, Content and VR went to defend the decision to only launch mainstream and performance products based on the Polaris architecture. “The reason Polaris is a big deal, is because I believe we will be able to grow that TAM [total addressable market] significantly.”

With Nvidia launching the GeForce GTX 1070 and 1080 for the Enthusiast market, and preparing its third Pascal chip – the GP106 – to fight Polaris 10/11, such approach might be a costly mistake. if the Radeon would be the only product family where the Polaris GPU architecture would make an appearance. In the same interview, Roy said something interesting: “We’re going on the record right now to say Polaris will expand the TAM. Full stop.”

At the same time, we managed to learn that SONY ran into a roadblock with their original PlayStation 4 plans. Just like all the previous consoles (PSX to PSOne, PS2, PS3), the plan was to re-do the silicon with a ‘simple’ die shrink, moving its APU and GPU combination from 28nm to 14nm. While this move was ‘easy’ in the past – you pay for the tapeout and NRE (Non-Recurring Engineering), neither Microsoft nor Sony were ready to pay for the cost of moving from a planar transistor (28nm) to a FinFET transistor design (14nm).

This ‘die-shrink’ requires to re-develop the same chip again, with a cost measured in excess of a hundred million dollars (est. $120-220 million). With Sony PlayStation VR retail packaging being a mess of cables and what appears to be a second video processing console, in the spring of 2014 SONY pulled the trigger and informed AMD that they would like to adopt AMD’s upcoming 14nm FinFET product line, based on successor of low-power Puma (16h) CPU and Polaris GPU processor architecture.

The only mandate the company received was to keep the hardware changes invisible to the game developers, but that was also changed when Polaris 10 delivered a substantial performance improvement over the original hardware. The new 14nm FinFET APU consists out of eight x86 LP cores at 2.1 GHz (they’re not Zen nor Jaguar) and a Polaris GPU, operating on 15-20% faster clock than the original PS4.

According to sources in the know, the Polaris for PlayStation Neo is clocked at 911 MHz, up from 800 MHz on the PS4. The number of units should increase from the current 1152. Apparently, we might see a number higher than 1500, and lower than 2560 cores which are physically packed inside the Polaris 10 GPU i.e. Radeon R9 400 Series. Still, the number of units is larger than Polaris 11 (Radeon R7 400 Series), and the memory controller is 256-bit wide, with GDDR5 memory running higher than the current 1.38 GHz QDR. Given the recent developments with 20nm GDDR5 modules, we should see a 1.75 GHz QDR, 7 Gbps clock – resulting in 224 GB/s, almost a 20% boost.

Internally known as PlayStation Neo, the console should make its debut at the Tokyo Game Show, with availability coming as soon as Holiday Season 2016 – in time for the PlayStation VR headset.

http://vrworld.com/2016/05/11/amd-confirms-sony-playstation-neo-based-zen-polaris/


PPS

 

[Assimilator]

Yes but, HBM2 is allooooooooooooooooooooooooooooooooooooooooooot faster than GDRR5X.

HBM1 -Fury X Memory bandwidth 512 (GB/s)
HBM2 - 1024 (GB/s) ?
GDDRX - 1080 Memory bandwidth 320 (GB/s)

Memory capacity is far more important than bandwidth. That's why Fury struggled against 980 Ti and AMD had to introduce the system memory cache, and why NVIDIA is sticking with GDDR5/X for everything except the highest-end cards.

 

[Eroticus]

Memory capacity is far more important than bandwidth. That's why Fury struggled against 980 Ti and AMD had to introduce the system memory cache, and why NVIDIA is sticking with GDDR5/X for everything except the highest-end cards.

Agree with you, but same story was with GDDR5 and GDRR3, Single cores and Multi cores CPUs.

New technology isn't working perfect at the start and this is normal, but some one brave should lead us to the light, and AMD always did that.

 

[FMinus]

Why is it so hard for people to understand that Enthusiast and High End cards make about of 5% of total GPU sales. That's a luxury market. Both AMD and Nvidia make a about 80% of their GPU profits on laptop, OEM, low-end and mid-range discreet cards.

This is what AMD is targeting with Polaris. It doesn't matter that it can't compete with the GTX 1080 or possibly the GTX 1070, it was never meant to go head to head with those two. But if one variant of the Polaris 10 chip can deliver performance north of R9 390x/R9 Fury for let's say $200-250 they have already won in that segment, and cut version of Polaris 10 might even go sub $200, and then there's Polaris 11 if they make discreet graphics out of it. Not to mention Apple is already going with Polaris GPUs in their future laptops.

Some tend to forget that people who buy cards for $500+ are in the vast minority. The market is blooming in the $100-$250 range.

Vega was always speculated to hit early Q1 of fiscal 2017, which starts for all companies pretty much in late 2016. It has nothing to do with Nvidia, and AMD starting to panic. It's proceeding as planned, and if they really moved it up, maybe the yields are better for both the chips and HBM2. Naturally AMD also knows that people want enthusiast grade cards, they have not forgot about them, so do AMD AIBs, even tho it doesn't make that much money in the big picture.

People are spelling doom and gloom, when this was never the case, not a single time in history. Both Nvidia and AMD are always neck on neck with their performance in comparable segments, I don't know why this should change now, except if they mess up real badly, but looking at the leaked benchmarks, Polaris is doing just fine for what it was meant to do, so really I don't know why people would worry at all.

Problem is people who don't follow this specific tech industry, don't know that Polaris was never meant as a top card, and they will be disappointed. But that happens all the time anyway, and the Nvidia "fanboys" are always louder, no wonder when they hold ~80% market share, it's to be expected. But at the end of the day, AMD Radeon Group is doing just fine, financial wise as performance wise with their past, current and future cards, no doubt about it.

 

[bug]

Why is it so hard for people to understand that Enthusiast and High End cards make about of 5% of total GPU sales. That's a luxury market. Both AMD and Nvidia make a about 80% of their GPU profits on laptop, OEM, low-end and mid-range discreet cards.

This is what AMD is targeting with Polaris. It doesn't matter that it can't compete with the GTX 1080 or possibly the GTX 1070, it was never meant to go head to head with those two. But if one variant of the Polaris 10 chip can deliver performance north of R9 390x/R9 Fury for let's say $200-250 they have already won in that segment, and cut version of Polaris 10 might even go sub $200, and then there's Polaris 11 if they make discreet graphics out of it. Not to mention Apple is already going with Polaris GPUs in their future laptops.

Bla, bla, bla. It's the halo products that make or break a brand, because they prove that you have the know how. Mercedes doesn't sell mostly S-Klasse, but can you imagine Mercedes not offering S-Klasse at all?
If AMD really had this planned all along, they would have mentioned it by now. Instead, they waited almost till the last moment and came up with the "expanding TAM for VR" that no one buys. Well, aside from fans, maybe.

 

[FMinus]

Bla, bla, bla. It's the halo products that make or break a brand, because they prove that you have the know how. Mercedes doesn't sell mostly S-Klasse, but can you imagine Mercedes not offering S-Klasse at all?
If AMD really had this planned all along, they would have mentioned it by now. Instead, they waited almost till the last moment and came up with the "expanding TAM for VR" that no one buys. Well, aside from fans, maybe.

Fury X is your S-class, not even a year old. AMD is telling you for about a year now that Polaris is a mid range card at best, just because you chose to ignore it or weren't in the loop doesn't mean this wasn't the plan all along.

I know it's hard to understand that most money comes from cheap(er) products even if you have it black and white on paper. Nobody aside of the internet fanboys gives a crap that brand A is 5-15 frames faster as brand B but both cost $700, exactly nobody. 80% PCs sold are in the OEM, Entry level or mid-range market, and let's not mention laptops.

Just because you want something and you judge a brand by that, doesn't mean it's a vital product for them. If AMD removed the Fury X, their profits would marginally decrease, same with nvidia and their Titan X and 980Ti. They sell nothing compared to GTX 950, 960.

 

[bug]

Fury X is your S-class, not even a year old. AMD is telling you for about a year now that Polaris is a mid range card at best, just because you chose to ignore it or weren't in the loop doesn't mean this wasn't the plan all along

I have honestly missed that. Do you have a year old link where AMD says Polaris will be mid-range only?

Just because you want something and you judge a brand by that, doesn't mean it's a vital product for them. If AMD removed the Fury X, their profits would marginally decrease, same with nvidia and their Titan X and 980Ti. They sell nothing compared to GTX 950, 960.

Their profits can't "marginally decrease", because they haven't been making any in years. However, if they gave up on high-end, the results would be dire, because the press would stop covering AMD. Look what happened when they stopped competing with intel at the top: everybody stopped caring about AMD CPU reviews. They all tell the same story: a recycled architecture, can't compete in perf/W with anything intel has, but they have a more capable IGP.

That is basically why AMD isn't doing what you claim they're doing. They're not dropping high-end, they're rushing Vega instead. It's just that probably after all these years in the red, they simply don't have the resources to go both after high-end and mid-range at the same time.

 

[EarthDog]

Google it....

 

[HumanSmoke]

Why now that GP104 seemed to gotten pushed-up form most original speculation (more end of July) is good, but this isn't?

There seem to be a ton of GTX 1080 cards in circulation. You would have to reserve judgement on Vega until the same could be said.

I don't think you could want/do two dissimilar memory controllers on one die.

You haven't been paying attention. AMD no longer puts together their own chips floorplans. That is now the province of Synopsys - who specialize in such things. It might not be trivial, but putting mix-and-match logic blocks together in a timely manner is what they do. It's the reason AMD partnered with them.

I find it hard to believe that AMD would venture a guess that Nvidia would be taking it easy. Myself (and hopefully most other people), I figured Nvidia's Pascal was going to give a great performance jump based on the available information floating around the internet.

AMD probably figured that Pascal wasn't going to be a big architectural leap and planned accordingly. What they weren't prepared for was the clock speed of the silicon - which is what gives Pascal the performance gain.

But GF promises...

Yeah, that's the problem. Globalfoundries always promise.
AMD are in court because Glofo promised that they had fixed their yield issues with 32nm
Globalfoundries, despite their rosy predictions for 28nm, had both ramp and yield issues.
Globalfoundies announce 20nmLPM among great fanfare and kill the process when everyone's back was turned
Remember when 14nm-XM was going to take the world by storm? GloFo doesn't remember.

Glofo is doing so well that the latest story doing the rounds could be a cut and paste from any of Glofo's previous exploits: AMD's Polaris fails clock speed validation at 850MHz
While AMD were quick to refute the claim they also claimed that Computex was going to see wall to wall Polaris. Guess we won't have long to wait to find out where the truth lies.

 

[Kanan]

Yes but, HBM2 is allooooooooooooooooooooooooooooooooooooooooooot faster than GDRR5X.

HBM1 -Fury X Memory bandwidth 512 (GB/s)
HBM2 - 1024 (GB/s) ?
GDDRX - 1080 Memory bandwidth 320 (GB/s)

HBM2 as seen on GP100 chips has 720 GB/s. Maybe not the highest version of it, but you never know.

GDDR5X speed is relative. 1) it's only 10 GHz GDDR5X used on GTX 1080, it gets as far as 14 GHz, 2) GTX 1080 is only using a 256 bit bus - other cards could use 384 bit or even 512 bit, which severely increases the bandwidth even with 10 GHz GDDR5X. HBM2 bandwidth isn't really greater then, but HBM2 consumes still less power, I guess.

 

[Caring1]

The title of this thread is still bugging me every time I see it
Must be my OCD, because when you "pull" something, it is cancelled, not moved forward.

 

[cadaveca]

should be "pull in" or "pull forward".

 

[Slizzo]

What they weren't prepared for was the clock speed of the silicon - which is what gives Pascal the performance gain.

I don't know, they probably could have seen that coming as it's a die shrink as well. Usually you can get clock speeds up if you have a better thermals. Also, Clocks aren't THAT much higher, GP104 is clocked at ~1600MHz on the GTX1080...

 

[cadaveca]

I don't know, they probably could have seen that coming as it's a die shrink as well. Usually you can get clock speeds up if you have a better thermals. Also, Clocks aren't THAT much higher, GP104 is clocked at ~1600MHz on the GTX1080...

But the demo had a card @ like 21000 MHz.... on stock cooling, apparently... You know, the Founder's edition....

https://www.techpowerup.com/222294/nvidia-geforce-gtx-1080-specifications-released

and the GPU barely scraped 67 °C under stress

 

[HumanSmoke]

I don't know, they probably could have seen that coming as it's a die shrink as well. Usually you can get clock speeds up if you have a better thermals. Also, Clocks aren't THAT much higher, GP104 is clocked at ~1600MHz on the GTX1080...

1600 was the base clock. How often during benchmarking would you expect the GPU to sit at its base clock? It's nominal boost is 1733MHz, and it's reported actual boost - if this is any indication- is between 1860MHz and 1886MHz. That speed represents a 57% higher clock than the 980Ti's 1202Mhz

 

[medi01]

I know it's hard to understand that most money comes from cheap(er) products

What you miss is "halo porducts" drive sales.
E.g. Fury release boosted 3xx sales, etc.

Whether AMD 470/480 would sell well, if there is nothing to take on at least 1070, uhm, who knows.
Will depend on reviewers really. If they focus on 'OMG< so slow" - nope. If on "OMG amazing value", possibly.

 

[Slizzo]

1600 was the base clock. How often during benchmarking would you expect the GPU to sit at its base clock? It's nominal boost is 1733MHz, and it's reported actual boost - if this is any indication- is between 1860MHz and 1886MHz. That speed represents a 57% higher clock than the 980Ti's 1202Mhz

I'm saying when taking die shrink into account it's not all that unheard of. Remember we skipped 20nm as well. It's likely that GloFo's 14nm will be able to hit the same clocks.

 

[bug]

I'm saying when taking die shrink into account it's not all that unheard of. Remember we skipped 20nm as well. It's likely that GloFo's 14nm will be able to hit the same clocks.

Actually, that's anything but likely.
It's a possibility, yes. But that possibility depends on the process itself being mature enough to mitigate leakage and stuff. If that's not up to par, you won't be able to jack the clock up. And at this point, we simply don't know. We know it's possible, but we're simply not sure.

 

[RejZoR]

I don't quite understand the point of GDDR5X considering it delivers LESS bandwidth on GTX 1080 than GTX 980Ti already has right now. Sure the bus can be narrower now, but dos that really affect the price all that much considering they are offsetting cheaper (narrower) bus with faster and more expensive memory which is hard to come by and is thus more expensive.