Question Speculation: RDNA3 + CDNA2 Architectures Thread

[uzzi38]

Man I have been dying to make this one for a while now.

First rumours for RDNA3 are here so new thread time!

Just going to start off with this one for now: kopite7kimi on Twitter: "@VideoCardz Ah, I mean a simple mcm design with 10240 cores is not enough. Because the lift from RDNA2 to RDNA3 is much bigger than from RDNA1 to RDNA2. We should expect many big improvements in GFX11. 🤔" / Twitter

 

[GodisanAtheist]

Not quite.
Both GPUs clocks pretty much the same, 7900 cache-memory speed and ROPs number overprovisioned.
Based on aggregate reviews data 7900xt is ~ 20% faster, which as you might guess, 4 additional CUs (5% more) can't account for. So the rest is basically IPC gain.

-Assuming the 7900XT is actually 20% faster than the 6950xt, nevertheless I doubt a 60CU N32 will be able to make up the 20% CU deficit vs a 6800xt even with a potential (very tenuous) 15% IPC increase.

Will likely be another 4060ti/7600 situation where it will barely tie in most scenarios and outright lose in others.

 

[insertcarehere]

Not quite.
Both GPUs clocks pretty much the same, 7900 cache-memory speed and ROPs number overprovisioned.
Based on aggregate reviews data 7900xt is ~ 20% faster, which as you might guess, 4 additional CUs (5% more) can't account for. So the rest is basically IPC gain.

They don't clock "pretty much the same" though, average game clocks for both have been tested through a variety of games:

	Techpowerup	Computerbase
6950XT	2404mhz	2392mhz
7900XT	2692mhz	2566mhz

Depending on the set of games tested that's an average of 7-11% clock speed increase right there.

And for the rest of the difference, we can't discount the extra memory bandwidth and ROPs contribution. As presumably AMD themselves had good reason to upsize these resources for the performance that AMD targeted..

 

[PJVol]

Depending on the set of games tested that's an average of 7-11% clock speed increase right there.

And for the rest of the difference, we can't discount the extra memory bandwidth and ROPs contribution. As presumably AMD themselves had good reason to upsize these resources for the performance that AMD targeted..

PCGH measured 2.4 for both, so it might be 100mhz difference indeed. Regarding the bandwidth or rops, is there any evidence that n31 or n21 lacking these resources to even start to consider?

 

[insertcarehere]

PCGH measured 2.4 for both, so it might be 100mhz difference indeed. Regarding the bandwidth or rops, is there any evidence that n31 or n21 lacking these resources to even start to consider?

Its not a question that we can test empirically (no way to significantly downclock N31 memory or disable ROPs).

But logically speaking, if AMD believed N21-esque memory/ROPs was fine for N31 performance levels to begin with (and they 99% would have simulated this config), then we'd see N31 launch with 4x MCDs and 128 ROPs. OEMs are not in the business of allocating extra die space for memory bandwidth (looks at Nvidia) and other resources if there's no corresponding performance increase.

 

[Timorous]

-Assuming the 7900XT is actually 20% faster than the 6950xt, nevertheless I doubt a 60CU N32 will be able to make up the 20% CU deficit vs a 6800xt even with a potential (very tenuous) 15% IPC increase.

Will likely be another 4060ti/7600 situation where it will barely tie in most scenarios and outright lose in others.

It needs to clock high without blowing the power budget and by high I mean 3Ghz +

RDNA 3 does seem to be able to clock high in N31, it just needs a ton of power to manage it so if AMD have been able to fix that issue with N32 then there is a chance.

 

[Heartbreaker]

It needs to clock high without blowing the power budget and by high I mean 3Ghz +

RDNA 3 does seem to be able to clock high in N31, it just needs a ton of power to manage it so if AMD have been able to fix that issue with N32 then there is a chance.

I don't why you keep coming back to this. 3GHz just doesn't seem to be on the table, not even NVidia on a more advanced process node is push that.

 

[PJVol]

Its not a question that we can test empirically (no way to significantly downclock N31 memory or disable ROPs).

Аctually it is, because overall performance is not just sum of the "bandwidth + ROPs + shaders + whatever-else" individually, so adding more bandwidth or ROPs won't give you performance gain right away (if at all) when let's say shader performance is limited by its own, or am I still unclear? Idk...

Either way, that wasn't the point of my post, and you might not read it entirely.
I could edit at that time, and it's not easy to keep up with native English speakers ))

nevertheless I doubt a 60CU N32 will be able to make up the 20% CU deficit vs a 6800xt even with a potential (very tenuous) 15% IPC increase

This I can't disagree with, though my prediction was a little more optimistic.

 

[Timorous]

I don't why you keep coming back to this. 3GHz just doesn't seem to be on the table, not even NVidia on a more advanced process node is push that.

As it stands, no, but we have seen high power use cards get tamed with a few tweaks. GTX480 GTX580 is such an example and if N32 is getting some fixes it will clock higher because there does not seem to be an architectural clock wall with N31, just the power use required gets stupid.

 

[Tuna-Fish]

I don't why you keep coming back to this. 3GHz just doesn't seem to be on the table, not even NVidia on a more advanced process node is push that.

N31 literally happily clocks >3GHz. Plenty of people have confirmed it. It just takes all the power to do that. This is likely a massive miss, as it costs transistors and power to shorten longest paths in a stage so that clocks can be pushed higher.

I am less sanguine than a lot of other people about AMD somehow "fixing" it in future steppings/dies. Generally, when you have clocking issues it's possible to massage some critical path to fix it. When you just burn too much power when you raise clocks, the only thing that will save you is a better process, and I don't think AMD is going to push Navi 32 to N4 or N3.

 

[Ajay]

I am less sanguine than a lot of other people about AMD somehow "fixing" it in future steppings/dies. Generally, when you have clocking issues it's possible to massage some critical path to fix it. When you just burn too much power when you raise clocks, the only thing that will save you is a better process, and I don't think AMD is going to push Navi 32 to N4 or N3.

Given where this generation of cards are at, it does seem to make zero sense to do a 're-spin', too much money for a small bump in sales. Reviews are already in for most of the cards folks will buy for gaming except for the N32 based cards. It's not very positive except for the 7900XTX, but that's cost more than the vast majority of games will spend. RDNA4, improvements in uArch and implementation - better use of the dual issue FP units, smaller process, etc. Hope AMD gets in right.

Nvidia did a bit better, save for the RAM issues (RT is awesome - don't have enough RAM, turn RT off ). Wackos.

 

[eek2121]

The idea that a 60/64CU N32 is going to perform like an 80cu 6900xt or even a 72CU 6800xt feels laughable at this point. We're basically looking at 6800 performance +5-10% at best from N32.

Probably why it's totally AWOL at this point, since the packaging costs of a Chiplet design at that tier are just making the economics funky, or N32 needs FSR 3 and framegen to pitch just like Ada.

Someone was able to fix the clocks at 1ghz for the 7600 and it was anywhere from 9%-22% faster.

I don't why you keep coming back to this. 3GHz just doesn't seem to be on the table, not even NVidia on a more advanced process node is push that.

Funny enough, OEMs are starting to release BIOS updates that allow for ~3 Ghz operation: https://videocardz.com/newz/asrock-...55-mhz-gpu-boost-with-the-new-extreme-oc-bios

I think AMD simply missed the thermal/power targets for RDNA3. A node shrink would probably do wonders here.

Regardless, the power limit required was “only” 480W. Apparently the cards pull 515W peak. That isn’t much higher than the 4090. This results in a 13.5% performance uplift.

If AMD had a proper air cooled design with this BIOS ready to go at launch, things likely would have turned out a bit different. This puts the card quite a bit closer to the 4090.

I have to wonder what a node shrink would do for RDNA 3.

 

[Tuna-Fish]

I reread my reply and felt I didn't put this clearly enough:

3GHz just doesn't seem to be on the table, not even NVidia on a more advanced process node is push that.

Clocks are not something where you can only have more if you are better or have a better process. They are an engineering tradeoff. The primary thing that determines highest achievable clock regardless of thermals, is the length of the longest critical path in your longest pipeline stage. It is, in principle, always possible to shorten it, and thus get higher peak clocks. The problem is that doing so has costs (mainly, increased transistor and power use of added latching and longer latencies for operations), and at some point those costs overwhelm the benefit of higher clocks. Looking at N31, what it really looks to me is that AMD designed a "speed demon", but the power consumption of the GPU at high speeds surprised them and caused performance to be less than desired.

This is a tale as old as time, see P4 and Bulldozer, just this time in GPU land. And as it was with those designs, unless TSMC pulls a rabbit out of a hat and saves AMD with a new, much less power-consuming process, likely the only way out of this is a substantial architectural rethink.

 

[Heartbreaker]

I reread my reply and felt I didn't put this clearly enough:

Clocks are not something where you can only have more if you are better or have a better process

Yes, but it's clear that isn't happening from the factory for RDNA 3 cards, and all the hopium in the world, won't change that.

 

[igor_kavinski]

If AMD had a proper air cooled design with this BIOS ready to go at launch, things likely would have turned out a bit different. This puts the card quite a bit closer to the 4090.

Probably what they plan to do with 7950 XTX, around the time when 4090 Ti is released. So about 1.5 years away?

 

[maddie]

I reread my reply and felt I didn't put this clearly enough:

Clocks are not something where you can only have more if you are better or have a better process. They are an engineering tradeoff. The primary thing that determines highest achievable clock regardless of thermals, is the length of the longest critical path in your longest pipeline stage. It is, in principle, always possible to shorten it, and thus get higher peak clocks. The problem is that doing so has costs (mainly, increased transistor and power use of added latching and longer latencies for operations), and at some point those costs overwhelm the benefit of higher clocks. Looking at N31, what it really looks to me is that AMD designed a "speed demon", but the power consumption of the GPU at high speeds surprised them and caused performance to be less than desired.

This is a tale as old as time, see P4 and Bulldozer, just this time in GPU land. And as it was with those designs, unless TSMC pulls a rabbit out of a hat and saves AMD with a new, much less power-consuming process, likely the only way out of this is a substantial architectural rethink.

Wonder what design tools they were using.

 

[maddie]

Probably what they plan to do with 7950 XTX, around the time when 4090 Ti is released. So about 1.5 years away?

That's RDNA4 time.

 

[igor_kavinski]

That's RDNA4 time.

That could always be delayed due to unforeseen issues and then we get RDNA3 refresh.

 

[Timorous]

N31 literally happily clocks >3GHz. Plenty of people have confirmed it. It just takes all the power to do that. This is likely a massive miss, as it costs transistors and power to shorten longest paths in a stage so that clocks can be pushed higher.

I am less sanguine than a lot of other people about AMD somehow "fixing" it in future steppings/dies. Generally, when you have clocking issues it's possible to massage some critical path to fix it. When you just burn too much power when you raise clocks, the only thing that will save you is a better process, and I don't think AMD is going to push Navi 32 to N4 or N3.

GF100 to GF110. GF110 has 10% higher clocks and 10% lower power use on the same node.

 

[Ajay]

That could always be delayed due to unforeseen issues and then we get RDNA3 refresh.

I keep hearing this, along with an NV refresh. I assume this is mainly binning better chips. Hopefully no delays, as this generation of GPUs is very meh.

 

[igor_kavinski]

Hopefully no delays, as this generation of GPUs is very meh.

Considering we still have a glut of RDNA2 GPUs and AMD themselves are putting RDNA2 in their market segment slide, It will be a miracle if RDNA4 launches on time.



I think judging by the lack of 6600 XT and 6900 XT in that slide, those are the ones running out soon at retailers.

 

[coercitiv]

Gamers Nexus has a video up discussing some new Acer cards that are still in the design phase. One of the cards is a dual 8-pin AMD card for which they are not allowed to mention the model. Considering the size of the cooling solution and the power delivery, I think there is a considerable chance this is a prototype for an N32 card. It could also be a 7900 XT, but in this case they would have little reason to hide the card model.

 

[igor_kavinski]

Hmmm...Acer getting ambitious. Seems their A770 BiFrost card sales were good enough for them to expand their AIB aspirations to include AMD GPUs too.

 

[jpiniero]

It could also be a 7900 XT, but in this case they would have little reason to hide the card model.

I will say maybe 7950 XTX. Faster memory and better binned for a bit more moola than the 7900 XTX. I don't think anything GDDR6 faster than 20 is shipping though.

 

[Aapje]

I keep hearing this, along with an NV refresh. I assume this is mainly binning better chips. Hopefully no delays, as this generation of GPUs is very meh.

Because of how AMD tends to both mess up their launch pricing and how long it takes them to fix their drivers, a refresh also has the benefit that they get new reviews with corrected prices and better drivers.

 

[Ajay]

Considering we still have a glut of RDNA2 GPUs and AMD themselves are putting RDNA2 in their market segment slide, It will be a miracle if RDNA4 launches on time.

View attachment 81106

I think judging by the lack of 6600 XT and 6900 XT in that slide, those are the ones running out soon at retailers.

Well, if Nvidia delays their next product release, then AMD can as well. Otherwise they have to compete. We know NV has cut wafer order, don't know about AMD. Given that no N32 GPU based AIB has been announced yet, they must indeed still have to much inventory in their '4K' range (and the prices on the 6950 XT surely indicated that). Nonetheless, development work has to continue.