News Into the GPU Chiplet Era: An interview with AMD's Sam Naffziger

[Admin]

AMD's Sam Naffziger talked with us about performance per watt and the GPU architectural journey of the past several generations, including RDNA 3 and what we can expect.

Into the GPU Chiplet Era: An interview with AMD's Sam Naffziger : Read more

 

[KananX]

Hopefully AMD will soundly take the performance crown this time, this will help balance the market and prevent Nvidia from demanding atrocious prices for their GPUs again. Also it’s time something changed, just with CPUs back then. Interesting article.

 

[rluker5]

Hopefully AMD will soundly take the performance crown this time, this will help balance the market and prevent Nvidia from demanding atrocious prices for their GPUs again. Also it’s time something changed, just with CPUs back then. Interesting article.

AMD prices their GPUs on the same price/perf curve as Nvidia. They are both willing participants in the high price duopoly. Your only hope for a disruption in this status quo is Intel GPUs.

 

[KananX]

AMD prices their GPUs on the same price/perf curve as Nvidia. They are both willing participants in the high price duopoly. Your only hope for a disruption in this status quo is Intel GPUs.

No, if AMD is faster, Nvidia can’t do another 3090 stunt with 1500-2000$ msrp. AMD already was cheaper this round, this will not change next round unless AMD is strictly faster and even then, because they are less popular you can expect lower prices than let’s say 3090. If Intel were my hope, I would be hopeless, their GPUs are a trainwreck.

 

[Makaveli]

AMD prices their GPUs on the same price/perf curve as Nvidia. They are both willing participants in the high price duopoly. Your only hope for a disruption in this status quo is Intel GPUs.

 

[KananX]

😂 it’s already working.

 

[waltc3]

Nice write up concerning RDNA3, which is naturally limited by what AMD wants to say about it right now, of course....😉 Back in 2020, when I first heard Wang discuss splitting off RDNA from CDNA architectures, I was pretty excited. It makes great sense to me. What I got from what he said was that for the gaming segment (RDNA) AMD was removing circuitry not especially tailored for gaming so that they'd have more room in each GPU to support more of the purely gaming circuitry and wouldn't have to devote space to circuitry that isn't as efficient or as robust for gaming as additional RDNA circuitry would be. Since CDNA is more targeted to compute-type processing, as opposed to game & texture acceleration, they could load up the CDNA architecture with more compute-type processing after having removed some of the non-essential gaming circuitry--whatever that may be. I thought it was brilliant and still do. There's definitely a split dynamic for the two types of GPUs presently. That's what I recall from Wang at that time, and it sounds like that's still the plan from your mention of the diverging circuitry in the article. Enjoyed it!...😉 Hard to think that we are not quite even half-way through 2022!...😉

 

[KananX]

Nice write up concerning RDNA3, which is naturally limited by what AMD wants to say about it right now, of course....😉 Back in 2020, when I first heard Wang discuss splitting off RDNA from CDNA architectures, I was pretty excited. It makes great sense to me. What I got from what he said was that for the gaming segment (RDNA) AMD was removing circuitry not especially tailored for gaming so that they'd have more room in each GPU to support more of the purely gaming circuitry and wouldn't have to devote space to circuitry that isn't as efficient or as robust for gaming as additional RDNA circuitry would be. Since CDNA is more targeted to compute-type processing, as opposed to game & texture acceleration, they could load up the CDNA architecture with more compute-type processing after having removed some of the non-essential gaming circuitry--whatever that may be. I thought it was brilliant and still do. There's definitely a split dynamic for the two types of GPUs presently. That's what I recall from Wang at that time, and it sounds like that's still the plan from your mention of the diverging circuitry in the article. Enjoyed it!...😉 Hard to think that we are not quite even half-way through 2022!...😉

Essentially AMD copied what Nvidia did with Maxwell back then. Meanwhile Nvidia themselves somewhat deviated from this by using typical things more meant for servers or workstation for gaming cards as well.

 

[spongiemaster]

AMD prices their GPUs on the same price/perf curve as Nvidia. They are both willing participants in the high price duopoly. Your only hope for a disruption in this status quo is Intel GPUs.

Exactly. AMD is no white knight. They want to make just as much money as Nvidia does. The problem is their fanbase won't pay Nvidia prices. Just look at the market now. AMD is offering a game bundle with all their cards and they are still mostly selling below MSRP. Not on the Nvidia side of things. AMD won't release their GPU's before Nvidia releases anything because, AMD wants to let Nvidia set the market. The more Nvidia charges, the more AMD can charge. There is no incentive for them to beat Nvidia to market as it will likely cost them money.

 

[escksu]

No, if AMD is faster, Nvidia can’t do another 3090 stunt with 1500-2000$ msrp. AMD already was cheaper this round, this will not change next round unless AMD is strictly faster and even then, because they are less popular you can expect lower prices than let’s say 3090. If Intel were my hope, I would be hopeless, their GPUs are a trainwreck.

Not going to work. This is because AMD is holding just around 17% of the market. It's around 1/5 that of Nvidia. Of course we have hpc sector but that's a different thing.

Btw, you won't see any price war in GPU nor CPU sector. Because there are literally just 2 companies (waiting for Intel to become 3rd) in the market, they will never enter a pricing competition.

What they all will do instead is to offer similar products at similar prices.

 

[KananX]

Not going to work. This is because AMD is holding just around 17% of the market. It's around 1/5 that of Nvidia. Of course we have hpc sector but that's a different thing.

Then you didn’t really follow what happened the last decades. Every time AMD was faster it pushed down Nvidias prices by a lot and also forced them to release faster GPUs for a lower price. Exactly what also happened with the 3080 last gen.

 

[escksu]

We need to implement some power cap or massive tax on these gaming cards. It ridiculous that so much power is used solely for personal gaming entertainment. It's really a waste of energy and resources. Pple are now gaming on phones, tablets and laptops. There are now even handheld "PC" which uses way less power. So why does a single gaming card need so much power??

 

[hannibal]

We need to implement some power cap or massive tax on these gaming cards. It ridiculous that so much power is used solely for personal gaming entertainment. It's really a waste of energy and resources. Pple are now gaming on phones, tablets and laptops. There are now even handheld "PC" which uses way less power. So why does a single gaming card need so much power??

Because we gamers demand more and more speed... And there really is competition in GPU market, so both companies push the limit as far as they can.
Same did happen in CPUs. Before AMD zen Intel cpus did not use much power at all. When AMD did get past Intel. Intel CPUs started using ridiculous amount of power to be competitive

 

[renz496]

No, if AMD is faster, Nvidia can’t do another 3090 stunt with 1500-2000$ msrp. AMD already was cheaper this round, this will not change next round unless AMD is strictly faster and even then, because they are less popular you can expect lower prices than let’s say 3090. If Intel were my hope, I would be hopeless, their GPUs are a trainwreck.

nvidia can still do it. Steve from GN said even if the 3090 command an absurd price people still have reasons to buy it: semi pro segment. even GN themselves are using 3090 for their productivity stuff rather than paying for true professional card like those RTX A6000. $1500 vs $5000 for the same exact performance except the pro card with the much needed certified drivers (which most often not needed for those in semi pro) and much more advance support. it is a segment that got ignored by AMD. also just because AMD is less popular they will not going to charge more. history says otherwise when AMD was first to launch 28nm GPU.

 

[TerryLaze]

When AMD did get past Intel. Intel CPUs started using ridiculous amount of power to be competitive

That's complete non-sense.
CPUs have always drawn ridiculous amounts of power under overclocking and the only thing that has changed is what intel considers overclocking, these new CPUs can handle a lot of power without melting so intel increased the amount of power that they will still warrant. (warrantee? )

The 5960x (8c/16t) back in 2014 was drawing the same 320W for 4.4Ghz that the 12900k draws for (8c/16t) at 5ghz + ,plus whatever the e-cores do.

Intel Core i7-7700K Kaby Lake Processor Review - Page 8

The Intel Core i7-6700K has become a firm favourite among PC enthusiasts despite the performance improvements being fairly slim compared to the previous generation, Intel hasn’t been under any…

www.eteknix.com

 

[KananX]

nvidia can still do it. Steve from GN said even if the 3090 command an absurd price people still have reasons to buy it: semi pro segment. even GN themselves are using 3090 for their productivity stuff rather than paying for true professional card like those RTX A6000. $1500 vs $5000 for the same exact performance except the pro card with the much needed certified drivers (which most often not needed for those in semi pro) and much more advance support. it is a segment that got ignored by AMD. also just because AMD is less popular they will not going to charge more. history says otherwise when AMD was first to launch 28nm GPU.

If you argue like that I will simply argue that the 3090 isn’t even a real GeForce then, it’s a Titan replacement semi pro card then which doesn’t compete against 6900 XT etc then, which makes my argument entirely valid again. Radeon Pro 6800 was also priced at 2000$ with 32 GB btw, so you’re wrong.

 

[renz496]

If you argue like that I will simply argue that the 3090 isn’t even a real GeForce then, it’s a Titan replacement semi pro card then which doesn’t compete against 6900 XT etc then, which makes my argument entirely valid again. Radeon Pro 6800 was also priced at 2000$ with 32 GB btw, so you’re wrong.

doesn't matter if they were geforce or not. the most important is they serve the semi pro market just like how nvidia want it to be. and nvidia compatibility with tons of professional software out there making some people value nvidia semi pro card even more than AMD true pro card. that's why nvidia can asking crazy price for the likes of 3090 even if price wise it might already competing with AMD true pro card.

 

[KananX]

doesn't matter if they were geforce or not. the most important is they serve the semi pro market just like how nvidia want it to be. and nvidia compatibility with tons of professional software out there making some people value nvidia semi pro card even more than AMD true pro card. that's why nvidia can asking crazy price for the likes of 3090 even if price wise it might already competing with AMD true pro card.

They just lose money on creators that would otherwise buy Quadro for a even higher price. If you think Nvidia makes high Bank with that, you’re wrong. Has a reason why they ditched true Titan, anyone who wants the full performance now in all apps has to ditch out for Quadro now, there’s still apps that run faster on Titan RTX than on 3090.

This is about CUDA advantage not about anything you said with market dominance. It’s purely software related. If AMD had good enough software alternative people would consider the 6900 XT instead.

 

[renz496]

They just lose money on creators that would otherwise buy Quadro for a even higher price.

nvidia simply admit the existence of semi pro market rather than ignoring it. plus for true professional work they still going to buy true pro card because they need those certified drivers and more advance support. things like Studio drivers? in the past only titan gets it. now even regular geforce have access to Studio drivers. meaning any geforce can be a semi pro card now. but for those need the fastest performance and big VRAM nvidia charging premium price with their x90 geforce.

It’s purely software related. If AMD had good enough software alternative people would consider the 6900 XT instead.

and people already considered that. for example i know there are interest by some people to run machine learning stuff on RDNA based card because the card have the hardware for it. but in the end they can't even start because AMD focusing ROCm support mostly for CDNA going forward.
the reality is while AMD try to catch up nvidia are not sitting still either. when AMD still try to make something like ROCm so people can migrate from CUDA nvidia already move beyond CUDA creating even more sophisticated ecosystem for their hardware. the end result it something like Optix where using the API end up being much faster than using CUDA itself.

 

[KananX]

nvidia simply admit the existence of semi pro market rather than ignoring it. plus for true professional work they still going to buy true pro card because they need those certified drivers and more advance support. things like Studio drivers? in the past only titan gets it. now even regular geforce have access to Studio drivers. meaning any geforce can be a semi pro card now. but for those need the fastest performance and big VRAM nvidia charging premium price with their x90 geforce.


and people already considered that. for example i know there are interest by some people to run machine learning stuff on RDNA based card because the card have the hardware for it. but in the end they can't even start because AMD focusing ROCm support mostly for CDNA going forward.
the reality is while AMD try to catch up nvidia are not sitting still either. when AMD still try to make something like ROCm so people can migrate from CUDA nvidia already move beyond CUDA creating even more sophisticated ecosystem for their hardware. the end result it something like Optix where using the API end up being much faster than using CUDA itself.

3090 is good but does not unlock full performance like Quadro does, just like I explained with Titan RTX earlier. It’s not a real Titan.

ROCm needs a few years to catch up still, RDNA2 support is coming as well, if it’s not already there. Now many are working to bring AMD up to speed because it’s way more widespread now with usage in many super computers. Also AMD is stocking up software engineers they are constantly hiring new ones. On the long run AMD will catch up with Nvidia just like they did with gaming.

 

[JarredWaltonGPU]

nvidia simply admit the existence of semi pro market rather than ignoring it. plus for true professional work they still going to buy true pro card because they need those certified drivers and more advance support. things like Studio drivers? in the past only titan gets it. now even regular geforce have access to Studio drivers. meaning any geforce can be a semi pro card now. but for those need the fastest performance and big VRAM nvidia charging premium price with their x90 geforce.

The Studio drivers are not the same as the old Titan drivers. The Titan drivers have improved support for a few pro applications. Just look at SPECviewperf. SNX, Creo, and Catia all run faster on the Titan RTX than even the 3090 Ti. The Nvidia Axxxx RTX GPUs blow away even the Titan in certain pro workloads.
https://www.tomshardware.com/review...professional-and-content-creation-performance

 

[JarredWaltonGPU]

ROCm needs a few years to catch up still, RDNA2 support is coming as well, if it’s not already there. Now many are working to bring AMD up to speed because it’s way more widespread now with usage in many super computers. Also AMD is stocking up software engineers they are constantly hiring new ones. On the long run AMD will catch up with Nvidia just like they did with gaming.

This is absolutely true. CUDA basically exists because of funding and support from the supercomputer sector. The governments put billions of dollars into research to make CUDA what it is. But now they're not pleased with being locked into CUDA (note that Nvidia is open sourcing its linux / data center drivers to combat this!), and now the same groups that helped Nvidia create CUDA are working with AMD to create ROCm. And it's being done in a far more open-source friendly way. "Fool me once..."

 

[KananX]

This is absolutely true. CUDA basically exists because of funding and support from the supercomputer sector. The governments put billions of dollars into research to make CUDA what it is. But now they're not pleased with being locked into CUDA (note that Nvidia is open sourcing its linux / data center drivers to combat this!), and now the same groups that helped Nvidia create CUDA are working with AMD to create ROCm. And it's being done in a far more open-source friendly way. "Fool me once..."

Yep, I think I read here that they simply want more alternatives, makes a ton of sense. Remember, AMD just was permitted to produce x86 way back then because IBM wanted more than just one company sourcing CPUs for their original IBM PCs. This led to AMD getting a license for x86 and now look how far this has come. Competition and alternatives are very important.

 

[renz496]

The Studio drivers are not the same as the old Titan drivers. The Titan drivers have improved support for a few pro applications. Just look at SPECviewperf. SNX, Creo, and Catia all run faster on the Titan RTX than even the 3090 Ti. The Nvidia Axxxx RTX GPUs blow away even the Titan in certain pro workloads.
https://www.tomshardware.com/review...professional-and-content-creation-performance

they are not the same but the intention still more or less the same: which is allowing non true pro nvidia card to have access to feature and performance are completely exclusive to pro cards only in the past even if they are not in full performance. part of this to encourage people to use their software and hardware. and hopefully in the future when they start doing more serious thing they will be buying nvidia true pro card.

 

[JayNor]

Intel's ARC GPUs with matrix operations will be supported in oneDNN. This is even on their entry level products. Looks like a good deal for students wanting to work on ai programs. Who's writing those type of reviews?