Discussion RDNA4 + CDNA3 Architectures Thread

DisEnchantment · Mar 23, 2022

With the GFX940 patches in full swing since first week of March, it is looking like MI300 is not far in the distant future!
Usually AMD takes around 3Qs to get the support in LLVM and amdgpu. Lately, since RDNA2 the window they push to add support for new devices is much reduced to prevent leaks.
But looking at the flurry of code in LLVM, it is a lot of commits. Maybe because US Govt is starting to prepare the SW environment for El Capitan (Maybe to avoid slow bring up situation like Frontier for example)

See here for the GFX940 specific commits

History for llvm/lib/Target/AMDGPU - llvm/llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. - History for llvm/lib/Target/AMDGPU - llvm/llvm-project

github.com

Or Phoronix

More AMD "GFX940" Enablement Work Landing In LLVM - Phoronix

www.phoronix.com

There is a lot more if you know whom to follow in LLVM review chains (before getting merged to github), but I am not going to link AMD employees.

I am starting to think MI300 will launch around the same time like Hopper probably only a couple of months later!
Although I believe Hopper had problems not having a host CPU capable of doing PCIe 5 in the very near future therefore it might have gotten pushed back a bit until SPR and Genoa arrives later in 2022.
If PVC slips again I believe MI300 could launch before it

This is nuts, MI100/200/300 cadence is impressive.

Previous thread on CDNA2 and RDNA3 here

Question - Speculation: RDNA3 + CDNA2 Architectures Thread

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...

forums.anandtech.com

TESKATLIPOKA · Apr 6, 2024

Abwx said:
You confused the GPU frequency with the BW, 2770 is the GPU frequency not the BW.
Same for the other GPU.

If you didn't notice he wrote 2770 for N48 but only 515 for N44.
Unless you want to claim It was a typo and he meant 2515 or 3515, where both are very unlikely for a good reason.
It's not likely for N44 to have lower or much higher clockspeed than N48.

Abwx · Apr 6, 2024

TESKATLIPOKA said:
If you didn't notice he wrote 2770 for N48 but only 515 for N44.
Unless you want to claim It was a typo I stand with IC BW.

I pointed it in my previous post, the 515 mean either 3515MHz or 2515MHz, the latter seems unexpectedly low, since it s clear that the BW is 288GB/s this number is forcibly related to the GPU frequency.

TESKATLIPOKA · Apr 6, 2024

Abwx said:
I pointed it in my previous post, the 515 mean either 3515MHz or 2515MHz, the latter seems unexpectedly low.

And precisely for that you are wrong. Look above
On the other hand for IC BW It makes much more sense, which is pretty similar to RDNA3 predecessors.

, the latter seems unexpectedly low, since it s clear that the BW is 288GB/s this number is forcibly related to the GPU frequency.

What forcibly related to GPU frequency? 288GB/s is precisely 128-bit 18gbps GDDR6. It has nothing to do with GPU frequency.

Timorous · Apr 6, 2024

Mahboi said:
32 CUs of non-broken RDNA 3/RDNA 3.5 would already probably handle 1440p quite well. The current broken 7600 is already very close. Assuming 20% extra perf, it would honestly surprise me that it can't reach 6700xt or so performance. And with the expected RDNA 4 improvements, maybe it'll be closer to 7700 xt performance...except it has that tiny bus.
Same with 64 CUs of non broken RDNA 3, that would already gain a solid 15% extra, it would put the 7800 xt around 4070 Ti performance? And then RDNA 4 improvements on top of it. I don't think there's going to be a lot of compromises compared to the current top range. Unless you think a 7900 xt is too weak for 4K.

7600XT and 7800XT already get + 15% from over clocking to around 3GHZ and a ram speed bump. If N44 and 48 clock closer to 3.3Ghz then N44 should land between 6750XT and 7700XT and N48 would be 7900XT level and that is with no IPC increase.

The XT is a tad weak for 4K ultra imo. it can do 60+ fine in a lot of games but there are also a fair few that fall below 60fps on average so you will need to use upscaling or tune settings to hit 60. Hence N48 with similar performance (and less bandwidth/cache to boot) will also have games where you need to compromise at 4K. If pricing is $600 or less then that is perfectly fine for that price point so not a big deal just how it will be.

N44 will be similar, it would offer 1440p 60 in a lot of titles but there will be a few where you need to turn down settings or use upscaling to hit that FPS figure. Again if pricing is $350 or less that is not really a problem.

So yea, I stand by what I said.

Mahboi · Apr 6, 2024

TESKATLIPOKA said:
That depends on what you mean by handle quite well. >60FPS?

In raster most likely yes, unless It's a game like City Skylines II.
RT is another story.

P.S. I am talking about N44

Yes, that's the numbers from RDNA 3. Which is broken. Hence why I said "non-broken", so you can expect a general flat 15% improvement. And there's RDNA 4's enhancements on top of that.

Timorous said:
The XT is a tad weak for 4K ultra imo.

Then we have very different appreciations of weak. I agree with everything else you said, I just don't think 55 or even 50 fps dips in a lot of games really matter.

Abwx · Apr 6, 2024

TESKATLIPOKA said:
And precisely for that you are wrong. Look above
On the other hand for IC BW It makes much more sense, which is pretty similar to RDNA3 predecessors.

The thing is that he said that the process is N4P, and looking at the perf/watt delta in respect of N5 and N6 this is perfectly in line with the stated frequencies in respect of a 7800XT and a 7600XT, that s exactly what to expect for 4 more CUs in the former case and same CU count for the latter while keeping the same TDPs at worse.

Timorous · Apr 6, 2024

Mahboi said:
Then we have very different appreciations of weak. I agree with everything else you said, I just don't think 55 or even 50 fps dips in a lot of games really matter.

Several are mid 40's or even lower though and don't even manage 60fps on average.

Alan wake 2, Cp2077, Jedi Survivor, lords of the fallen, last of us part 1, starfield etc.

By the time N48 and N44 launch that list will grow hence why I think N44 will be a 1080p card primarily and N48 will be a 1440p card primarily. Not saying you can't use them at higher resolutions because plenty of games will be fine but enough new AAA games won't be that I don't think they can advertise them as 1440p/4k cards with a straight face. OTOH AMD lied through their teeth in the RDNA 3 reveal so maybe they just do it anyway.

Timorous · Apr 6, 2024

Abwx said:
The thing is that he said that the process is N4P, and looking at the perf/watt delta in respect of N5 and N6 this is perfectly in line with the stated frequencies in respect of a 7800XT and a 7600XT, that s exactly what to expect for 4 more CUs in the former case and same CU count for the latter while keeping the same TDPs at worse.

That assumes the same v/f curve which if so means AMD will have two duds on the trot. Not impossible but it would be a disaster for the client GPU market.

TESKATLIPOKA · Apr 6, 2024

Abwx said:
The thing is that he said that the process is N4P, and looking at the perf/watt delta in respect of N5 adn N6 this is perfectlly in line with the stated frequencies in respect of a 7800xt and a 7600XT, that s exactly what to expect for 4 more CUs in the former case and same CU count for the latter.

N32: 2430MHz 263W -> N48: 2770MHz (+340MHz)
Believable for comparable or lower TDP.

N33: 2655-2755MHz 165-190W -> N44: 2515-3515MHz
Either 140-240MHz lower clock than N33 using N6 process or 760-860MHz higher clock, which is 28-32% higher than N33.
Both are extremes regardless of how I look at It.
Ok, 3515MHz could be "believable" that AMD "fixed It" + better process, but memory BW should have been improved by a lot, but wasn't at all.
Then N48 also shouldn't have only 2770MHz, that's 745MHz difference between N44 vs N48! Then I would rather believe It's actually only 2515MHz but with much lower TDP.

If you want to believe that It's clockspeed + Typo, feel free to do so, but stop saying that I am wrong when you have no proof, It's just your speculation without even knowing the actual TDP.

Abwx · Apr 6, 2024

Timorous said:
That assumes the same v/f curve which if so means AMD will have two duds on the trot. Not impossible but it would be a disaster for the client GPU market.

Of course, that s very conservative, i assumed that nothing was fixed in respect of the v/f curve, so we can take this as a worst case that is unlikely to happen if they made some progress in this register.

Abwx · Apr 6, 2024

TESKATLIPOKA said:
Then N48 also shouldn't have only 2770MHz, that's 745MHz difference between N44 vs N48! Then I would rather believe It's actually only 2515MHz but with much lower TDP.

If you want to believe that It's clockspeed + Typo, feel free to do so, but stop saying that I am wrong when you have no proof, It's just your speculation without even knowing the actual TDP.

That s because the 32 CUs GPU goes from N6 to N4P while the 64 CUs start from N5, not counting that there s 4 more CUs, the smaller GPU has forcibly much more margin to increase frequency.

Edit : 2515MHz would mean that it would barely perform better than the 7600, and likely lower than the 7600XT, that just doesnt make sense.

7600XT clock up to 2760 and from N6 to N4P there s about 22% higher frequency at same power, that get us close to 3.4GHz, and that s assuming that the v/f curve is unchanged.

jpiniero · Apr 6, 2024

Abwx said:
7600XT clock up to 2760 and from N6 to N4P there s about 22% higher frequency at same power, that get us close to 3.4GHz, and that s assuming that the v/f curve is unchanged.

I don't think it works that way...

Ghostsonplanets · Apr 6, 2024

https://twitter.com/x/status/1776553577928069464

The Navi 48 samples are capable of 50 TFLOPS on a ~215W board.

jpiniero · Apr 6, 2024

Ghostsonplanets said:
https://twitter.com/x/status/1776553577928069464

The Navi 48 samples are capable of 50 TFLOPS on a ~215W board.

Maybe that's how they are improving RT... it has more compute power compared to bandwidth/raster performance.

Tuna-Fish · Apr 6, 2024

Assuming 32 wgp, 50TF is ~3.06GHz.

igor_kavinski · Apr 6, 2024

Ghostsonplanets said:
The Navi 48 samples are capable of 50 TFLOPS on a ~215W board.

That would be SWEET!

CouncilorIrissa · Apr 6, 2024

Sane power consumption even? Amazing.

Tuna-Fish · Apr 6, 2024

Tbh even if the chip is capable of that, I'd fully expect AMD to juice the XT model to the very limit. Then turning on the eco mode is lef for the user.

SolidQ · Apr 6, 2024

Ghostsonplanets said:
on a ~215W board.

looks good. 4070ti is 28W

50 TFLOPS

7900XT is 51.5, but would interesting how much RDNA3 1TF compared to RDNA4 TF

CouncilorIrissa · Apr 6, 2024

Tuna-Fish said:
Tbh even if the chip is capable of that, I'd fully expect AMD to juice the XT model to the very limit. Then turning on the eco mode is lef for the user.

I wish this BS was left to OC editions from AIBs.

Bring back sane power consumption and non-aircraft carrier-sized cooling setups.

Ghostsonplanets · Apr 6, 2024

50 TFLOPs for a 32 WGP design with board power of 215W is a quite solid PPW improvement over the 7900 XT. Really showing off RDNA 4 impressive PPA improvements over RDNA 3, as AMD is greatly reducing area and power usage while exhibiting same performance as a bigger RDNA 3 card. Definitely feels like we're bound for another RDNA 2 generation of greats cards. Shame that only two dies remained though.

I now wonder what will be the board power of Navi 44. N44 is basically half of N48 and will probably clock a bit lower too. Maybe something like 100W board power? Regardless, AMD has good designs for both DIY and also to be slotted into Gaming Laptops alongside Strix Halo.

Timorous · Apr 6, 2024

Ghostsonplanets said:
https://twitter.com/x/status/1776553577928069464

The Navi 48 samples are capable of 50 TFLOPS on a ~215W board.

So 3.05 Ghz then which is not bad at 215W board power.

Performance of 64CUs at 3Ghz with enough memory bandwidth should be in the ballpark of an OC'd 7900GRE and on the heels of the 7900XT. Maybe more if there is an IPC improvement or the sustained clock in games is a bit higher.

Suspect N44 will clock higher and target 120-130W something like that. Should be good for 25-30 Tflops which will be a bit below the 7700XT unless there is an IPC improvement as well.

CouncilorIrissa · Apr 6, 2024

Timorous said:
So 3.05 Ghz then which is not bad at 215W board power.

Performance of 64CUs at 3Ghz with enough memory bandwidth should be in the ballpark of an OC'd 7900GRE and on the heels of the 7900XT. Maybe more if there is an IPC improvement or the sustained clock in games is a bit higher.

Suspect N44 will clock higher and target 120-130W something like that. Should be good for 25-30 Tflops which will be a bit below the 7700XT unless there is an IPC improvement as well.

N44 still only has 2/3 of 7700XT's memory bandwidth, so little chance of competing against the 7700XT (unless it's somehow massively less membw hungry)

SolidQ · Apr 6, 2024

CouncilorIrissa said:
so little chance of competing against the 7700XT

Even if RTX 3070 level with 16GB, would be nice for 300$

PJVol · Apr 6, 2024

Interesting, is this the same shady TFlops count as with the Rdna 3?

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Platinum Member

Lifer

Platinum Member

Golden Member

Senior member

Lifer

Golden Member

Golden Member

Platinum Member

Lifer

Lifer

Lifer

Senior member

Lifer

Golden Member

Lifer

Senior member

Golden Member

Senior member

Senior member

Senior member

Golden Member

Senior member

Senior member

Senior member