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]

They probably aimed at 3GHz clocks, but could not reach them because of power consumption. Also, being the GCD die small but getting the most of the power consumption, that means that thermal density on the GCD is quite high.

Zen4 has higher thermal density than N31, so that shouldn't be such a problem.
N24 has 107W TBP. N21 at those clocks and ported to 6nm could be ~450W TBP.
It will be interesting to see how high N31 will clock at a lot higher TBP.

 

[Shmee]

I am very curious to see how high end partner cards such as Sapphire Nitro will perform, especially if they have 3 8 pins and higher clocks.

 

[fleshconsumed]

I am very curious to see how high end partner cards such as Sapphire Nitro will perform, especially if they have 3 8 pins and higher clocks.

Performance wise it's going to be diminishing returns - you may get 3-5% higher performance at expense of 15-20% higher power consumption. The biggest appeal of aftermarket cards is lower temps and quieter cooling. I would gladly pay $20-50 extra for a better quieter cooler, unfortunately AIBs typically charge way more than that.

 

[Shmee]

Hmm, I would expect more performance than that, especially if one tweaks the voltage and clocks more themselves. Unless you think that the GPU is already close to the limit? Or maybe it is starved for faster memory? I guess time and reviews will tell.

 

[GodisanAtheist]

I literally thought so when I read this the day before they advanced together )

View attachment 70411

- Its all shadows on walls cast by a flickering flame, but it does make one wonder if NV pushed their 4090 out the door without adequate QC and clocked to the gills because they spooked at what AMD could potentially bring to the table.

 

[maddie]

Performance wise it's going to be diminishing returns - you may get 3-5% higher performance at expense of 15-20% higher power consumption. The biggest appeal of aftermarket cards is lower temps and quieter cooling. I would gladly pay $20-50 extra for a better quieter cooler, unfortunately AIBs typically charge way more than that.

What part of the F/V curve is the 7900XTX? We just don't know enough yet to be so certain.

 

[AnandThenMan]

Maybe I'm reading things wrong but are some of you suggesting AMD should have the same performance with far less power draw? Sure Nvidia is the performance king but even putting pricing aside, power draw and form factor has become absurd.

 

[Karnak]

Someone on the internet made these graphs:
View attachment 70419

Why are ppl making graphs with avg. fps (TPU) vs. max. fps (AMDs "up to")? That's just completely pointless.

IMO it doesn't bode well that AMD is only providing such vague "up to" numbers instead of exact fps figures like they did two years ago with RDNA2 vs. Ampere. Are they not confident enough and do they know that Reviews with avg. and min. fps will show less improvements over RDNA2?

 

[Timorous]

Why are ppl making graphs with avg. fps (TPU) vs. max. fps (AMDs "up to")? That's just completely pointless.

IMO it doesn't bode well that AMD is only providing such vague "up to" numbers instead of exact fps figures like they did two years ago with RDNA2 vs. Ampere. Are they not confident enough and do they know that Reviews with avg. and min. fps will show less improvements over RDNA2?

They used UPTO in the RDNA2 presentation as well. It is just legal CYA speak to show that is the 'upto' average when using a certain config.

They are not maximum framerates.

See - upto even though they were the averages and were pretty much spot on when independently reviewed.

 

[fleshconsumed]

What part of the F/V curve is the 7900XTX? We just don't know enough yet to be so certain.

No we don't, but both AMD/nvidia have been clocking their designs to the tilt in the past 5 years, it is not going to change with ADA/RDNA3.

And then somebody also posted this:

https://twitter.com/x/status/1588329053530251265

 

[Joe NYC]

There were not really any good upgrades from Intel for HEDT for nearly a decade (on the CPU side).

I think a good upgrade is 2x the old component performance and it may not be until maybe Zen 4 V-Cache in single core performance.

 

[JayMX]

If "only 3% OC" what is the point of 3x8 power sockets and reinforced cooling? IMHO this information doesn't make much sense.

 

[Karnak]

They used UPTO in the RDNA2 presentation as well. It is just legal CYA speak to show that is the 'upto' average when using a certain config.

They are not maximum framerates.

See - upto even though they were the averages and were pretty much spot on when independently reviewed.

Missed that one. My bad. Thought they were actual numbers without that little thing at the top.

Althoug this time there are still less games they provided numbers for. I mean especially without using FSR performance mode which is not... that great if you need to back up your numbers with that one. In CP it's even worse because they have to use FSR 1.0 which is just... garbage like DLSS1.

They spent more time talking about extremely high refresh rates and fps due to DP2.1 (don't know if it's actually true, but I feel like it is) rather than "real" 4K benchmarks and RT which I don't like in terms of what we'll get in 5 weeks.

 

[Timorous]

Missed that one. My bad. Thought they were actual numbers without that little thing at the top.

Althoug this time there are still less games they provided numbers for. I mean especially without using FSR performance mode which is not... that great if you need to back up your numbers with that one. In CP it's even worse because they have to use FSR 1.0 which is just... garbage like DLSS1.

They spent more time talking about extremely high refresh rates and fps due to DP2.1 (don't know if it's actually true, but I feel like it is) rather than "real" 4K benchmarks and RT which I don't like in terms of what we'll get in 5 weeks.

In the top chart they show 5 games. This go around they did 3 raster and 3 RT games with a similar format. they did not give an average, perhaps because they were mixing raster and RT and perhaps because 3 games is a bit few to do it with but that RDNA2 slide is very accurate vs reality on average so I don't think AMD would have picked 3 titles that were huge outliers for the raster uptick. They are probably 3 games in the middleish of the distribution. In fact if I am doing that kind of graph and I am AMD I am skewing slightly lower than the middle of the distribution so there can be a positive surprise when reviews release. AMD did that with the 5800X3D so they do have form but lets wait and see how it pans out.

In any even using a 1.54x scaling factor from the 6950XT feels ballpark correct. The 3 games shown in the reveal average out to 1.57x raster as well so that also fits and I think that is the ballpark. Some sites might be closer to 1.6x others might be closer to 1.5x and when 3D centre do the meta analysis I think 1.55 give or take is a reasonable expectation.

 

[PJVol]

Maybe I'm reading things wrong but are some of you suggesting AMD should have the same performance with far less power draw? Sure Nvidia is the performance king but even putting pricing aside, power draw and form factor has become absurd.

It's amazing how so many people still compare 7900XTX to the higher tier 4090. Even AMD officials (Azor) have made it clear that the XTX is a competitor to the 4080, aside from the obvious things you mentioned above.
Frank Azor on perf. comparisons

Also didn't get what percentage for "one-touch" (Auto) OC he assumed at 11:55 ?

 

[A///]

It's amazing how so many people still compare 7900XTX to the higher tier 4090. Even AMD officials (Azor) have made it clear that the XTX is a competitor to the 4080.
Frank Azor on perf. comparisons

Also didn't catch what percentage for "one-touch" OC assumed at 11:55.

I saw that earlier today. Hmm do we think AMD is holding back a 7950 or 7970 card that competes with the 4090? ill wait for benchmarks but as it stands now the amd card is a winner compared to whatever sorry excuse nvidia is pushing out as a 4080 with 16 gb that doubles as a fire hazard.

 

[gdansk]

I don't believe there is any real 4090 competitor planned.
I don't think we'll see either 2xGCD or a bigger GCD until RDNA4.
There might be a stacked cache variant planned for a 2023 refresh. But it won't change the relative position of things.

 

[Det0x]

Someone on the internet made these graphs:
View attachment 70419

And some more from the interwebz:

https://www.reddit.com/r/Amd/comments/ymc4aw

 

[A///]

that makes sense. I didn't think there'd be many flagships around. The xt and xtx make compelling purchases for those considering a 4080.

 

[Karnak]

It's amazing how so many people still compare 7900XTX to the higher tier 4090. Even AMD officials (Azor) have made it clear that the XTX is a competitor to the 4080,

If AMDs benchmarks are true the 7900XTX is like 15-20% faster in raster. Plus you get an additional +50% VRAM. Doesn't look 'just' like a 4080 competitor to me. Heck, it's even $200 cheaper and AD103 is not that much bigger than the N31 GCD.

Sure, when it comes to RT the 4080 will wipe the floor with the 7900XTX especially in RT heavy games, but from what perspective is it a 4080 competitor? Just doesn't add up to me but maybe there's something I am missing.

 

[A///]

I think a good upgrade is 2x the old component performance and it may not be until maybe Zen 4 V-Cache in single core performance.

you should only consider cpu and gpu upgrades every other generation.

 

[RnR_au]

No we don't, but both AMD/nvidia have been clocking their designs to the tilt in the past 5 years, it is not going to change with ADA/RDNA3.

And then somebody also posted this:

https://twitter.com/x/status/1588329053530251265

As mentioned earlier, that preliminary driver could be power limited. So the coretek tweet could be correct but newer drivers will unlock greater overclocks.

We just don't know where on the v/f curve AMD placed the reference models. I hope there is some meat left on the bones for the AIBs.

 

[Hail The Brain Slug]

As mentioned earlier, that preliminary driver could be power limited. So the coretek tweet could be correct but newer drivers will unlock greater overclocks.

If the driver or bios is hard-locked to the pci-e spec for maximum power, it would be limited to 366W (66W 12V from slot + 2x150W 8 pin). Guess what 3% more power than 355W is?

Edit: before anybody mentions about pcie slot being 75W, the spec only allows 5.5A max on 12V, which is 66W

 

[Shmee]

If the driver or bios is hard-locked to the pci-e spec for maximum power, it would be limited to 366W (66W 12V from slot + 2x150W 8 pin). Guess what 3% more power than 355W is?

Edit: before anybody mentions about pcie slot being 75W, the spec only allows 5.5A max on 12V, which is 66W
View attachment 70441

Hmm, that may be it, that the reference cards are BIOS locked to spec for maximum power draw. This should change on models with 3 8pin connectors.

 

[Hail The Brain Slug]

Hmm, that may be it, that the reference cards are BIOS locked to spec for maximum power draw. This should change on models with 3 8pin connectors.

That's not exactly true. The 6950XT reference card is a 335W TDP and looking at the bios, it allows +20% power limit which puts it over 400W, beyond spec for 2x8 pin + slot. I assume the preliminary driver or bios they are using for the 7900XTX is hard locked to spec for one reason or another.

I expect the final release cards to offer some kind of power headroom, even if only 10-15%. Hopefully MorePowerTool can be updated to work on RDNA3.