AMD Retreating from Enthusiast Graphics Segment with RDNA4?

[ViperXZ]

Or they could not have spun off GF & settled with Intel out of court for a puny billion back in 2008/09 huh? The FX was a disaster in major parts because AMD was lagging Intel 1-2.5 nodes behind. We can speculate all we want but it is what it is ~

Nah, the FX was a disaster because it was a terrible design. The node was the smallest of its problems, since it was a huge chip with high clocks anyway.

 

[R0H1T]

And who's to say what it's performance could've been on 14nm? There was also cat cores at the time which were killed by Intel's contra revenues ~ again good potential but not efficient enough. Should we go on?

AMD Beema/Mullins Architecture & Performance Preview

www.anandtech.com

 

[ViperXZ]

You're just speculating, so am I. ATI saved AMD, the same money they put into ATI to buy it, couldve saved them in other ways. A very very simple concept, you don't want to grasp.

 

[R0H1T]

The speculation part was started by you. We can do this forever btw

 

[Colddecked]

Clearly AMD needs to higher the both of you as they have no clue what they’re doing.

/s

In a freely ideal situation, sure. Our assumptions on a business’s decisions have no relevance to what’s actually going on as we have no real information on why those decisions are made. The number of times I’ve experienced a client telling me to redesign something because it’s “easy” or better when they have no experience, education, or factual information as to why and how continues to grow and is an ever useful reminder that as much as we want to believe something is simple from the outside, it’s not.

Must all disagreements lead to "Oh you so smart why don't they hire you"? This is a discussion amongst enthusiasts with different opinions. Can you not try to cut off the discussion by going that route?

Obviously they have their calculations and projections and spending that kind of money on GPUs isn't a priority. WE ALL AGREE ON THIS POINT. For a company who's stock price doubled since 2020, that has made multi billion dollar accquistions since 2020, I just happen to think they should have invested more in securing a more advanced node for Navi 3X.

 

[ViperXZ]

Obviously they have their calculations and projections and spending that kind of money on GPUs isn't a priority. WE ALL AGREE ON THIS POINT. For a company who's stock price doubled since 2020, that has made multi billion dollar accquistions since 2020, I just happen to think they should have invested more in securing a more advanced node for Navi 3X.

And lets not forget that its a huge achievement to be able to compete with Intel and Nvidia at the same time - while being way smaller than both.

The speculation part was started by you. We can do this forever btw

I don't have a problem with it, since I'm not offended easily in discussions (like many other people regularly are on the internet).

 

[rv8000]

this is nonsense, as it's not about "clue", if you would have properly understood what I was saying, it is about "possibilities" and "money" and "opportunities".

In other words: AMD does know everything I said. And more

What a foolish statement. AMD has internal departments and reports to make these decisions and what makes the most fiscal sense.

Believe your wild conclusions all you’d like, but until you run a multi billion dollar with multiple departments providing you with the background on how and why these decisions are made, you nor I have zero ground to stand on.

Seeing as how well AMD as a whole has done since the changes of CEO to Lisa, initial design and improvements of Zen, and multiplying of net worth, it seems they’re well off with the decisions they’ve made.

 

[ViperXZ]

What a foolish statement. AMD has internal departments and reports to make these decisions and what makes the most fiscal sense.

Believe your wild conclusions all you’d like, but until you run a multi billion dollar with multiple departments providing you with the background on how and why these decisions are made, you nor I have zero ground to stand on.

Seeing as how well AMD as a whole has done since the changes of CEO to Lisa, initial design and improvements of Zen, and multiplying of net worth, it seems they’re well off with the decisions they’ve made.

And still you didn't understand the point of my posts, not at all. It's beyond comical what you are doing

It's funny to be offended for a company, which i did not attack (I'm a huge AMD fan btw)

 

[Colddecked]

And still you didn't understand the point of my posts, not at all. It's beyond comical what you are doing

It's funny to be offended for a company, which i did not attack (I'm a huge AMD fan btw)

I'm sure Lisa is crying in the corner of her office after reading our comments/criticisms

 

[R0H1T]

And lets not forget that its a huge achievement to be able to compete with Intel and Nvidia at the same time - while being way smaller than both.


I don't have a problem with it, since I'm not offended easily in discussions (like many other people regularly are on the internet).

There's nothing to be offended by any way, still not sure what your point is? As far as I'm concerned AMD wouldn't be alive & kicking if it weren't for ATI, at the same time ATI (standalobe) would probably be bankrupt as well if they didn't merge with AMD. Remember PowerVR, & where are they now?

Imagination and Apple Renew IP Licensing Agreement

www.anandtech.com

Imagination Technologies to Be Acquired by Canyon Bridge for £550M; MIPS Sold to Tallwood VC

www.anandtech.com

Basically a one trick pony with Apple's graphics IP deal.

 

[ViperXZ]

There's nothing to be offended by any way, still not sure what your point is? As far as I'm concerned AMD wouldn't be alive & kicking if it weren't for ATI, at the same time ATI (standalobe) would probably be bankrupt as well if they didn't merge with AMD. Remember PowerVR, & where are they now?

Imagination and Apple Renew IP Licensing Agreement

www.anandtech.com

You know that over a billion dollar can do a lot to save you in other ways, through other designs for example? Your point is weak at best. You "being sure" doesn't mean much, since you have no proof for a alternate future. Your speculation isn't better than mine.

I'm sure Lisa is crying in the corner of her office after reading our comments/criticisms

I bet she loves these discussions and would never see a negative thing about it. It's very possible they had these kind of discussions and then settled on the MCM design, because monolithic wasn't in the books.

 

[rv8000]

And still you didn't understand the point of my posts, not at all. It's beyond comical what you are doing

It's funny to be offended for a company, which i did not attack (I'm a huge AMD fan btw)

Im not offended, you’re simply making assumptions as to why a massive company makes the decisions they do without any means to ground overly idealistic ideas.

Everyone is moving towards chiplet design, AMD looks to have been playing the long game, and has arguably made the smart move.

 

[R0H1T]

You know that over a billion dollar can do a lot to save you in other ways, through other designs for example? Your point is weak at best. You "being sure" doesn't mean much, since you have no proof for a alternate future. Your speculation isn't better than mine.

Yes it is unless we're also going to speculate that ATI could've spawned an AI empire as big as Nvidia. There's really only 3-5 major IC makers who can afford to compete on the bleeding edge all the time ~ Apple, Nvidia, QC & of course AMD/Intel & that's because it's stupidly expensive to make these top chips & you need tons of volumes to recoup your investment! ATI couldn't afford that with their limited range of products at the time, any way happy speculating

 

[Colddecked]

It's very possible they had these kind of discussions and then settled on the MCM design, because monolithic wasn't in the books.

It was a bold decision, much like going chiplets for Ryzen was. I can appreciate that. But unlike Ryzen, the timing wasn't as perfect and competition not as stagnant. Nvidia learns from mistakes, like not jumping on advanced nodes for turing and ampere. They've since corrected that mistake and likely won't make it again.

 

[ViperXZ]

It was a bold decision, much like going chiplets for Ryzen was. I can appreciate that. But unlike Ryzen, the timing wasn't as perfect and competition not as stagnant. Nvidia learns from mistakes, like not jumping on advanced nodes for turing and ampere. They've since corrected that mistake and likely won't make it again.

Agreeing that MCM is better at all for AMD (performance wise), that it's not just to save wafers in 5nm as half of the chip is produced on 6nm basically, I would say that they still didn't use the possibility to produce a simply bigger chip - especially if you consider that with MCM it's easier possible than with monolithic.

That being said, latency is about 10% worse than with a monolithic RDNA 3, in this case 7600: https://chipsandcheese.com/2023/06/14/latency-testing-is-hard-rdna-3-power-saving/

So, I don't see a upside in MCM aside from, again, saving wafers, and that is only so needed for AMD because they share it with their CPUs - something which would not be the case if Radeon would still be ATI.

Nvidia will go for CHIPLET as soon as it makes SENSE for them, as in, better performance or efficiency. Both things which AMD did not achieve this time.

 

[rv8000]

Agreeing that MCM is better at all for AMD (performance wise), that it's not just to save wafers in 5nm as half of the chip is produced on 6nm basically, I would say that they still didn't use the possibility to produce a simply bigger chip - especially if you consider that with MCM it's easier possible than with monolithic.

That being said, latency is about 10% worse than with a monolithic RDNA 3, in this case 7600: https://chipsandcheese.com/2023/06/14/latency-testing-is-hard-rdna-3-power-saving/

So, I don't see a upside in MCM aside from, again, saving wafers, and that is only so needed for AMD because they share it with their CPUs - something which would not be the case if Radeon would still be ATI.

Nvidia will go for CHIPLET as soon as it makes SENSE for them, as in, better performance or efficiency. Both things which AMD did not achieve this time.

How do you not see an upside in MCM? Higher yields, lower costs (that aren’t necessarily passed on to the consumer), both a performance and efficiency improvement (see below 6900XT vs 7900XTX), you’re ignoring facts easily accessible to you. Not to mention density and cooling requirements. Also AMD, unlike nvidia, didn’t increase their msrp across every tier of GPU; again we know their are cost savings associated with MCM.

Interconnects are the obvious hurdle, and AMD has proven at this point via Zen evolution improvements can and will be made. I have no doubt Nvidia can easily go down this route.

 

[R0H1T]

Another major upside is dedicated accelerators like for RT or AI, which can be connected over separate chiplets & should save a lot of power/heat when they're not in use. This would also allow the "GCD" to possibly clock higher. The downsides are generally kept in check with massive caches, the bigger issue would be data transfer which for now is bottlenecked by IF on AMD's side.

 

[ViperXZ]

How do you not see an upside in MCM? Higher yields, lower costs (that aren’t necessarily passed on to the consumer), both a performance and efficiency improvement (see below 6900XT vs 7900XTX), you’re ignoring facts easily accessible to you. Not to mention density and cooling requirements. Also AMD, unlike nvidia, didn’t increase their msrp across every tier of GPU; again we know their are cost savings associated with MCM.

View attachment 308378View attachment 308380Interconnects are the obvious hurdle, and AMD has proven at this point via Zen evolution improvements can and will be made. I have no doubt Nvidia can easily go down this route.

MCM is only always to save costs, unless you can pair CPU CHIPLETS as with Ryzen to increase overall core count, which did not happen with the GCDs used on Radeon, as did not happen any big increases in "core counts" (aka stream processors). In other words, in order for AMD to make PERFORMANCE use of MCM they wouldve needed to pair multiple GCDs, so that it works as a single GPU, with no big downsides - something which was speculated from users for years, but unfortunately AMD did not deliver this, instead chosing to do a MCM design with just 1 GCD and MCDs - IE a 5+6 nm hybrid approach to save costs.

That being said, my arguments are MONOLITHIC vs MCM. MCM = only a upside in performance if you use multiple CORE CHIPLETS, not just 1x CORE CHIPLET + MCDs.

The INTER DIE COMMUNICATION increases power (=decreases efficiency) because of the longer traveling time from GCD to MCD, which is not a thing in a monolithic design.

 

[rv8000]

MCM is only always to save costs, unless you can pair CPU CHIPLETS as with Ryzen to increase overall core count, which did not happen with the GCDs used on Radeon, as did not happen any big increases in "core counts" (aka stream processors). In other words, in order for AMD to make PERFORMANCE use of MCM they wouldve needed to pair multiple GCDs, so that it works as a single GPU, with no big downsides - something which was speculated from users for years, but unfortunately AMD did not deliver this, instead chosing to do a MCM design with just 1 GCD and MCDs - IE a 5+6 nm hybrid approach to save costs.

That being said, my arguments are MONOLITHIC vs MCM. MCM = only a upside in performance if you use multiple CORE CHIPLETS, not just 1x CORE CHIPLET + MCDs.

The INTER DIE COMMUNICATION increases power (=decreases efficiency) because of the longer traveling time from GCD to MCD, which is not a thing in a monolithic design.

A few simple questions:

Does the 7900 XTX not out perform the 6900XT?

Does the 7900 XTX have better, on average and across the board, better efficiency than the 6900XT?

I’ll spoil it for you, yes and yes.

You can even look at Zen4 vs gen13 intel, in terms of efficiency intel gets slaughtered. And then we can look at the effects of die shrinks and transistor density making chips ever more heat dense and difficult to cool. There are way more benefits to MCM than costs savings with the ever encroaching limitations of silicon. To say otherwise is to be naive.

Leave the engineering to Intel, AMD and Nvidia, your armchair serves little purpose.

 

[Colddecked]

A few simple questions:

Does the 7900 XTX not out perform the 6900XT?

Does the 7900 XTX have better, on average and across the board, better efficiency than the 6900XT?

I’ll spoil it for you, yes and yes.

You can even look at Zen4 vs gen13 intel, in terms of efficiency intel gets slaughtered. And then we can look at the effects of die shrinks and transistor density making chips ever more heat dense and difficult to cool. There are way more benefits to MCM than costs savings with the ever encroaching limitations of silicon. To say otherwise is to be naive.

Leave the engineering to Intel, AMD and Nvidia, your armchair serves little purpose.

7900xtx has a number of advantages over 6900xt. Faster ram, wider bus, newer process and architecture. The question is would 7900xtx perform better if it was monolithic built on 5nm and the answer is without a doubt yes. It just would have been more expensive.

 

[ViperXZ]

7900xtx has a number of advantages over 6900xt. Faster ram, wider bus, newer process and architecture. The question is would 7900xtx perform better if it was monolithic built on 5nm and the answer is without a doubt yes. It just would have been more expensive.

More expensive for AMD + less GPUs delivered, because it eats the 5nm wafers then, and it does not get more 5nm wafers - Radeon has a set amount of wafers, they will not get any more just because the chip is bigger. We saw this downside with Big Navi having huge issues being available at release and even later. AMD did not want this again, aside from cost, which you rightly described, so they did this MCM.

Fact is, with monolithic it's realistic to expect that it would've also had a higher clock due to higher efficiency + better latency, which is also higher perf. This would've made it easily faster than the 4080, but with the downside that AMD would not have made big financial gains with it. Then realistic to expect, it would've been priced maybe at 1200 instead of 1000, due to being faster than 4080, despite other short comings (no DLSS 3, no other NV features). Could've been still very profitable, but = less chips, less graphics cards = less money.

 

[rv8000]

7900xtx has a number of advantages over 6900xt. Faster ram, wider bus, newer process and architecture. The question is would 7900xtx perform better if it was monolithic built on 5nm and the answer is without a doubt yes. It just would have been more expensive.

Maybe.

You’re ignoring what I was replying to, and making a point of.

Viper stated as some self proclaimed fact that the only advantage is “saving dies” (whatever that actually means), and that Navi31 offers no energy efficiency or performance improvements over Navi21, which is objectively false. To what those performance and efficiency improvements are directly attributed to we cannot say for sure, not without some white paper documentation dissecting architectural improvement.

As it exists, the 7900XTX is a faster, more efficient GPU, that is a fact. We have no proof of what a monolithic navi31 would have been in terms of hardware specs and how scaling/die size limitations may have benefited or harmed the design and performance.

Which all circles back to Viper just stating a bunch of bs assumptions without any way to ever prove, let alone understand why AMD chose to approach the design of Navi31 the way they did.

If I were to make some sort of equivalent statement along the thought process of both vipers and yours, I should be calling AMD incapable of making good decisions for not making a monolithic GPU with 12k shaders, 64GB of ram, on a 512bit bus with 384mb of infinity cache on a 2nm process. Do you see how that’s ridiculous?

It’s one thing to discuss hypothetical technology, but something else entirely to call out a company for not doing something when you have no idea of what the engineering limitations or problems that design could entail.

 

[Colddecked]

Maybe.

You’re ignoring what I was replying to, and making a point of.

If I were to make some sort of equivalent statement along the thought process of both vipers and yours, I should be calling AMD incapable of making good decisions for not making a monolithic GPU with 12k shaders, 64GB of ram, on a 512bit bus with 384mb of infinity cache on a 2nm process. Do you see how that’s ridiculous?

It’s one thing to discuss hypothetical technology, but something else entirely to call out a company for not doing something when you have no idea of what the engineering limitations or problems that design could entail.

And you are leaving out 7900xtx being on a smaller process explains why it's more efficient than 6900xt. It has very little to do with being a chiplet. As we can see 4090 is monolithic, is more efficient and has better performance (thanks in no small part to... being on a smaller node).

Absolutely no one is saying oh if it was monolithic it should have 12k cores, 64gb HBM3 ram, etc. Show me where I even imply that. Quit with the straw man stuff THAT'S ridiculous.

Viper is saying MCM make more sense when you know, have actually more than 1 (graphics) chiplet. How is this not sinking in? If you don't believe that you are drowning in AMD kool aid brother.

Yes AMD got a head start on MCM for GPUs, which should I hope help them in the future since yes chiplets are probably the way forward. But their GPU situation right now is as dire as its ever been. Not saying they're doomed but the optimism is down for me.

 

[rv8000]

And you are leaving out 7900xtx being on a smaller process explains why it's more efficient than 6900xt. It has very little to do with being a chiplet. As we can see 4090 is monolithic, is more efficient and has better performance (thanks in no small part to... being on a smaller node).

Absolutely no one is saying oh if it was monolithic it should have 12k cores, 64gb HBM3 ram, etc. Show me where I even imply that. Quit with the straw man stuff THAT'S ridiculous.

Viper is saying MCM make more sense when you know, have actually more than 1 (graphics) chiplet. How is this not sinking in? If you don't believe that you are drowning in AMD kool aid brother.

Yes AMD got a head start on MCM for GPUs, which should I hope help them in the future since yes chiplets are probably the way forward. But their GPU situation right now is as dire as it’s ever been. Not saying they're doomed but the optimism is down for me.

Take a breath and read what I wrote. I never said efficiency or performance improvements were exclusive to MCM. Seriously… take a breath and read.

While you both continue to falsely claim you know exactly what the minutia of Navi31s design attributes to x% performance increase, I’m simply stating what Navi31 is based on the information we have.

The crystal ball armchair engineer logic is strong with you two; we have no clue what a monolithic die Navi31 would be for better or worse. I’m just not naive enough to believe I know better than the 100s of engineers working for these companies.

 

[Colddecked]

How do you not see an upside in MCM? Higher yields, lower costs (that aren’t necessarily passed on to the consumer), both a performance and efficiency improvement (see below 6900XT vs 7900XTX), you’re ignoring facts easily accessible to you. Not to mention density and cooling requirements. Also AMD, unlike nvidia, didn’t increase their msrp across every tier of GPU; again we know their are cost savings associated with MCM.

Interconnects are the obvious hurdle, and AMD has proven at this point via Zen evolution improvements can and will be made. I have no doubt Nvidia can easily go down this route.

I did read what you wrote. You are saying the upside to MCM is better performance and efficiency, without mentioning process and architecture improvements which are more important. Almost everything positive you attribute to MCM is more likely due to being on 5nm and being a new architecture. Convenient you used 4k data to show performance difference between 6900xt and 7900xtx where bus size and memory speed difference are more likely bigger factors than being MCM.

Also that in bold is not the truth, they have increased MSRP across most every tiers. I am wrong about 7900xtx and 6900xt MSRP.