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

 

[TESKATLIPOKA]

The one part I'm still waiting on is AMD explaining where the promised >50% perf/watt increase went. Some reviews like Computerbase even reported worse perf/watt gen over gen in some cases.

That 54% improvement was at 300W, not sure in what game.
TPU measured 34.5% better efficiency than RX 6900XT in Cyberpunk, so that's not that bad.
ComputerBase at 4K measured 41% higher FPS in Doom Eternal and 20% higher power draw in the same game, adding to 17.5% improvement. Which is not very good.

They got worse efficiency than 6900XT with 144 FPS limiter, true.
If you check out what is in these graphs, then you will know the reason.


Limit to 144 FPS causes 34% loss in FPS(217->144), but power draw was lowered by just by 9W(354W -> 345W), so no wonder efficiency got much worse.

 

[TESKATLIPOKA]

RDNA3 uses software scheduler like Nvidia compared to hardware one in RDNA2, right? Software + AMD is not the best combination.
I think we will see pretty good improvements with the next big Adrenaline update.

 

[biostud]

RDNA3 uses software scheduler like Nvidia compared to hardware one in RDNA2, right? Software + AMD is not the best combination.
I think we will see pretty good improvements with the next big Adrenaline update.

So the hopes are that the driver team can improve the scheduler over time? Maybe around the time N32 launches?

 

[TESKATLIPOKA]

So the hopes are that the driver team can improve the scheduler over time? Maybe around the time N32 launches?

It doesn't mean RDNA3 will suddenly crush Ada or anything like that, but I think with this change, It will be true more than ever that AMD GPUs will get better over time.

At Computerbase they compared PPC(perf. per count) between RX 7900XT(2.2GHz) and RX6900XT(2.3GHz). I made a table based on that.

Raster	RX 6900XT 2.3 GHz	RX 7900XT 2.2 GHz	Raster + RT	RX 6900XT 2.3 GHz	RX 7900XT 2.2 GHz
Call of Duty: Vanguard	100 %	117 %	Cyberpunk 2077	100 %	127 %
Cyberpunk 2077	100 %	116 %	Deathloop	100 %	111 %
Deathloop	100 %	113 %	Doom Eternal	100 %	132 %
Doom Eternal	100 %	107 %	Dying Light 2	100 %	118 %
Dying Light 2	100 %	104 %	F1 22	100 %	116 %
F1 22	100 %	106 %	Far Cry 6	100 %	108 %
Far Cry 6	100 %	110 %	Ghostwire Tokyo	100 %	113 %
Forza Horizon 5	100 %	106 %	Metro Exodus Enhanced	100 %	122 %
Ghostwire Tokyo	100 %	101 %	Saints Row	100 %	111 %
God of War	100 %	123 %	Spider-Man Remastered	100 %	125 %
Halo Infinite	100 %	109 %
Saints Row	100 %	107 %
Shadow Warrior 3	100 %	100 %
Sniper Elite 5	100 %	116 %
Spider-Man Remastered	100 %	122 %
Tiny Tina's Wonderlands	100 %	105 %
Average	100 %	109 %	Average	100 %	118 %

In 2 games we can see >20% improvement even without RT enabled, but in 2 games there is pretty much none.
I think in games with <10% difference we will see some improvement later.

 

[RnR_au]

The one part I'm still waiting on is AMD explaining where the promised >50% perf/watt increase went. Some reviews like Computerbase even reported worse perf/watt gen over gen in some cases.

The sign should be the other way. AMD's claim on this matter;

AMD RDNA 3 architecture delivers up to 54% more performance per watt

From https://www.amd.com/en/press-releas...d-s-most-advanced-gaming-graphics-cards-built

So its <54% perf/watt increase depending on the game. And at TBP of 300W according to the footnotes.

 

[moinmoin]

The sign should be the other way.

Now it most certainly should be. But originally it wasn't:

 

[maddie]

Now it most certainly should be. But originally it wasn't:

I guess the difference between everything working and everything working as planned.

 

[Mopetar]

I guess the difference between everything working and everything working as planned.

In theory everything should work out in practice, but in practice things rarely work out according to theory.

 

[Joe NYC]

JayzTwoCents is overclocking RDNA 3 card.

Some strange results. He is able to overclock GCD > 3 GHz, 3.2 GHz in fact, but as he does, the GCD uses up most of the power limits and downclocks memory.

So, it seems, with higher power limits (everything else being the same), there could be some performance gains.

 

[Topweasel]

JayzTwoCents is overclocking RDNA 3 card.

Some strange results. He is able to overclock GCD > 3 GHz, 3.2 GHz in fact, but as he does, the GCD uses up most of the power limits and downclocks memory.

So, it seems, with higher power limits (everything else being the same), there could be some performance gains.

The question will be whether AMD will allow AIB's to build for higher power limits and whether they even thought to include them considering the second rate design AMD products get until they prove their performance. If they do then the AMD reference cards will be half way decent ITX cards, but those AIB cards will look mighty good. But people hitting 3GHz should shut up the "power bug, second silicon guys", though we might need to see a second round of AIB's before we see some really good cards.

 

[uzzi38]

The question will be whether AMD will allow AIB's to build for higher power limits and whether they even thought to include them considering the second rate design AMD products get until they prove their performance. If they do then the AMD reference cards will be half way decent ITX cards, but those AIB cards will look mighty good. But people hitting 3GHz should shut up the "power bug, second silicon guys", though we might need to see a second round of AIB's before we see some really good cards.

There'll eventually be MPT updates that will let you go beyond what AIBs allow you to set.

Just a matter of time.

Also you should see the clocks these things can hit with some tuning in non-gaming workloads, it's mental.

 

[leoneazzurro]

There'll eventually be MPT updates that will let you go beyond what AIBs allow you to set.

Just a matter of time.

Also you should see the clocks these things can hit with some tuning in non-gaming workloads, it's mental.

It would be interesting to understand why they can't hit these clocks in games. Compute tasks work in the shaders' ALUs anyway, so it seems clear that some other part of the pipeline is involved in clock reduction due to increased power consumption.

 

[TESKATLIPOKA]

From All The Watts!!

So the N33 is no longer 77** and 76** but a tier lower.
Does It mean N32 wil be 78** and 77**, or will they go up to 79**?

 

[jpiniero]

From All The Watts!!
View attachment 73193
So the N33 is no longer 77** and 76** but a tier lower.
Does It mean N32 wil be 78** and 77**, or will they go up to 79**?

That's mobile. Maybe they will do N31 mobile?

 

[leoneazzurro]

From All The Watts!!
View attachment 73193
So the N33 is no longer 77** and 76** but a tier lower.
Does It mean N32 wil be 78** and 77**, or will they go up to 79**?

Rumors were always that way (N32 being 7800/7700 and N33 being 7600/7500). Clocks were wrong but configurations (especially from Angostronomics) seem consistent. 7800 should be full N32 with 256 bit memory interface while 7700 should be a cut down N32 with 192 bit interface. Or are you speaking about mobile variants? In that case, it's possible we get only the 7800M variant (with N32 but cut down and/or with reduced clocks). Last gen we had no 6900M and Nvidia 4090 mobile should be a 4080 chip, in reality. .

 

[TESKATLIPOKA]

Rumors were always that way (N32 being 7800/7700 and N33 being 7600/7500). Clocks were wrong but configurations (especially from Angostronomics) seem consistent. 7800 should be full N32 with 256 bit memory interface while 7700 should be a cut down N32 with 192 bit interface.

All The Watts!! was talking about 7700/7600 for N33, but he could have been wrong. It's still unknown, If he is not really greymon55.
If the full N32 will be only 7800M then It looks like AMD won't have anything to compete with 4090M based on ADA103.
I don't think we will see N31 in mobile.
I wonder If they will rename 6500M to 7400M or will keep the old name.

 

[leoneazzurro]

I don't think that either, power envelope and memory bus of N31 are outside the usual mobile footprint and even AD103 will have severe power restrictions. In mobile space, power is the most limiting factor.

 

[jpiniero]

I don't think that either, power envelope and memory bus of N31 are outside the usual mobile footprint and even AD103 will have severe power restrictions. In mobile space, power is the most limiting factor.

I believe nVidia is going up to 180 W for the 4090 Mobile.

 

[TESKATLIPOKA]

I don't think that either, power envelope and memory bus of N31 are outside the usual mobile footprint and even AD103 will have severe power restrictions. In mobile space, power is the most limiting factor.

AD103 should be still better than N31.
It looks like Nvidia is better at very limiting power settings, at least based on desktop cards.

Spoiler: V-Sync

• V-Sync: If you don't need the highest framerate and want to conserve power, running at 60 FPS is a good option. In this test, we run Cyberpunk 2077 at 1920x1080, capped to 60 FPS. This test is also useful in testing a graphic card's ability to react to situations with only low power requirements. For graphics cards that can't reach 60 FPS at 1080p, we report the power draw at the highest achievable frame rate

 

[leoneazzurro]

AD103 should be still better than N31.
It looks like Nvidia is better at very limiting power settings, at least based on desktop cards.

Spoiler: V-Sync

• V-Sync: If you don't need the highest framerate and want to conserve power, running at 60 FPS is a good option. In this test, we run Cyberpunk 2077 at 1920x1080, capped to 60 FPS. This test is also useful in testing a graphic card's ability to react to situations with only low power requirements. For graphics cards that can't reach 60 FPS at 1080p, we report the power draw at the highest achievable frame rate

View attachment 73195

Yes, that was foreseeable due to the result of the desktop cards, also some difference is due to the different memory and bus size. N32 should fare better anyway and possibly new drivers could also give a boost, so in power limited scenarios Ad103 will be faster anyway, but the difference could be less dramatic than expected.

 

[Glo.]

I wonder If they will rename 6500M to 7400M or will keep the old name.

Why would they?

 

[leoneazzurro]

I believe nVidia is going up to 180 W for the 4090 Mobile.

Considering that top 3080 mobile SKU already can be set to 175W, it's not surprising.

 

[Glo.]

To be honest, I expected that N32 would be cut down to 24 CUs, and 6 GB VRAM, not 28 CUs and 6 GB VRAM.

To be honest guys, I will be VERY happy if on desktop we will get 28 CUs, 6 GB VRAM, and low clocks, and no 6 pin connector for 7500 series GPU.

 

[TESKATLIPOKA]

Why would they?

Because of new generation? Or will they discontinue It? Not like It made a good name with all the compromises made just to make It as small as possible.

To be honest, I expected that N32 would be cut down to 24 CUs, and 6 GB VRAM, not 28 CUs and 6 GB VRAM.

To be honest guys, I will be VERY happy if on desktop we will get 28 CUs, 6 GB VRAM, and low clocks, and no 6 pin connector for 7500 series GPU.

Then it would have a lot weaker performance and AMD would need to significantly reduce the price. They won't do It in my opinion If that chip can be set to higher TDP which will increase performance along with the selling price.

 

[Glo.]

Because of new generation? Or will they discontinue It? Not like It made a good name with all the compromises made just to make It as small as possible.

Then it would have a lot weaker performance and AMD would need to significantly reduce the price. They won't do It in my opinion If that chip can be set to higher TDP which will increase performance along with the selling price.

If TDP for mobile parts is 120W with AIB 6600 XT clock levels, then we could be easily be able to get over 2 GHz, with 96 bit bus clock speeds for 7500 series SKU and no 6 pin connector.

That would be vastly faster than RX 6500 XT, while consuming less power, and delivering more VRAM.