Discussion RDNA4 + CDNA3 Architectures Thread

[DisEnchantment]

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

 

[MrTeal]

The Amazon UK Best Sellers has the NVidia GT 710, as the #6 best seller.

Granted, it's super cheap, but exactly how old would your IGPU have to be, to actually be worse than that?

There's also a market for people running AM4 CPUs and needing the cheapest thing with current driver support to get a display output.

 

[MarutTheSlayer]

That's because TAA is crap, not because DLSS makes it magically better than native.

lol. DLSS is great at antialiasing based on ML. Take the very best raster based AA and put it up against DLAA, the difference will be clear as day.

 

[Thunder 57]

lol. DLSS is great at antialiasing based on ML. Take the very best raster based AA and put it up against DLAA, the difference will be clear as day.

OK now you start by saying DLSS but say to compare it to DLAA. I'd agree with you on DLAA, but not DLSS being "better than native". Even then I'd only say that with current crappy AA we are forced to use because of deferred rendering. Compared to MSAA or SSAA, not so sure I'd agree, But this is an RDNA4 thread so I'll leave it there.

 

[MarutTheSlayer]

OK now you start by saying DLSS but say to compare it to DLAA. I'd agree with you on DLAA, but not DLSS being "better than native". Even then I'd only say that with current crappy AA we are forced to use because of deferred rendering. Compared to MSAA or SSAA, not so sure I'd agree, But this is an RDNA4 thread so I'll leave it there.

That's because DLAA native at 1440P is what is carried forward to DLSS Q at 4K, DLSS simply upscales based on a lower native render resolution with applied DLAA, which is 1440P for 4K. That's why it looks great.

 

[Saylick]

You can find it if you look for it on YT in the last couple of days. Could easily have been fudged though but interesting buildup.

There's a ton of faked benchmarks on YT. I wouldn't trust it.

 

[Heartbreaker]

It's probably bcos of FSR4 not being ready and reviewers giving it a rough time when DLSS4 is ready to roll, and brings substantial image quality improvements to 40 series as well.

Need to work on 4X+ Frame Gen, and those Frame aren't going to fake themselves...

 

[eek2121]

Good lord AMD needs to put this farce of a launch out of its misery.

Physical cards are at physical retailers and we still have essentially *nothing*.

It isn’t a ‘farce of a launch’

They previewed at CES. Board partners do have the chips, but not the proper drivers. They haven’t launched yet mostly because they are finalizing drivers. A few leaks have dropped, but AMD did not provide anything beyond enough functionality to allow board makers to test.

If you want to know where it lands, it is around 7900XTX in terms of raster, sometimes behind, sometimes ahead. RT will be faster. Release price is supposed to be $479 for the XT model. There seems to be confusion if the XT model launches this month, the non-XT, or both, but something is.

Don’t believe me? that is fine, I cited no sources. Wait for the launch.

 

[Saylick]

It isn’t a ‘farce of a launch’

They previewed at CES. Board partners do have the chips, but not the proper drivers. They haven’t launched yet mostly because they are finalizing drivers. A few leaks have dropped, but AMD did not provide anything beyond enough functionality to allow board makers to test.

If you want to know where it lands, it is around 7900XTX in terms of raster, sometimes behind, sometimes ahead. RT will be faster. Release price is supposed to be $479 for the XT model. There seems to be confusion if the XT model launches this month, the non-XT, or both, but something is.

Don’t believe me? that is fine, I cited no sources. Wait for the launch.

That's solid performance for a solid price then, since it would be faster than a 5070 and cheaper, while also offering more VRAM. FSR 4 would be icing on the cake, assuming it is comparable to DLSS 2. If so, welcome back, AMD.

 

[Thunder 57]

It isn’t a ‘farce of a launch’

They previewed at CES. Board partners do have the chips, but not the proper drivers. They haven’t launched yet mostly because they are finalizing drivers. A few leaks have dropped, but AMD did not provide anything beyond enough functionality to allow board makers to test.

If you want to know where it lands, it is around 7900XTX in terms of raster, sometimes behind, sometimes ahead. RT will be faster. Release price is supposed to be $479 for the XT model. There seems to be confusion if the XT model launches this month, the non-XT, or both, but something is.

Don’t believe me? that is fine, I cited no sources. Wait for the launch.

If what you say is true the 5070 is DoA, which it should be regardless because of its crippled VRAM.

 

[GTracing]

That's solid performance for a solid price then, since it would be faster than a 5070 and cheaper, while also offering more VRAM. FSR 4 would be icing on the cake, assuming it is comparable to DLSS 2. If so, welcome back, AMD.

The 7900 XTX is 25-30% faster than the 4070 TI. Nvidia is claiming a ~20% improvement for the 5070 TI over the 4070 TI (in Resident Evil 4 with ray tracing on). If the 9070 XT is as fast as a 7900 XTX in raster, then it will beat the 5070 TI.

That would also mean that AMD has improved the IPC (for lack of a better term) by ~30% with RDNA4.

An improvement like that isn't unprecedented, but I'll believe it when I see it.

 

[SolidQ]

One month ago benchmarks. Currently might be difference

 

[techjunkie123]

That would also mean that AMD has improved the IPC (for lack of a better term) by ~30% with RDNA4.

No? 20% better clocks and 5% IPC?

 

[techjunkie123]

One month ago benchmarks. Currently might be difference

Bruh if it's this performance and 479$, they're gonna sell like hotcakes.

 

[SolidQ]

vs 4070S/4080

That 304W 9070XT. I'd assuming final results will be better

 

[gdansk]

Bruh if it's this performance and 479$, they're gonna sell like hotcakes.

It's not $479 if that's accurate. AMD would never.

 

[SolidQ]

It's not $479 if that's accurate.

Yeah, and that 304W + from december tests. That mean final perfomance gonna be higher. My bet maybe 600$? 150$ less than 5070ti

 

[MarutTheSlayer]

It's not $479 if that's accurate. AMD would never.

Well, Lisa can do the most funny thing on stage if it is true.
Drop a price of $749, and just say kidding as the numbers rearrange themselves for $479. Then a new 🐐 is on stage to take over from 1080 Ti, and will earn Madam Su a ton of boss memes.

 

[adroc_thurston]

No? 20% better clocks and 5% IPC?

yes. per-CU perf is waaaaay up.

I'd assuming final results will be better

yea, it's exactly ~7900XTX perf now.

It's not $479 if that's accurate. AMD would never.

ding-ding-ding, with Blackwell being what it is (poo), AMD can price whatever. $549 is the optimistic pick.

 

[MarutTheSlayer]

yea, it's exactly ~7900XTX perf now.

Do you know if the power goes up to 300W for the reference card? Retail specs show 260 odd watts for the XT, and a couple of YT videos of the 9070 XT supposedly (could be fake too) show performance like the XTX but power sits between 235 to 265W.

 

[Kepler_L2]

yes. per-CU perf is waaaaay up.

RDNA3 bros we are so back

 

[adroc_thurston]

Do you know if the power goes up to 300W for the reference card?

Genuinely no idea what the current MBA power is.
But 40W a large difference do not make anyway.

RDNA3 bros we are so back

this time it's NV missing out on fmax bumps (or at least feels like it). woo.
if only KircheisNavi40 was alive

 

[MarutTheSlayer]

Genuinely no idea what the current MBA power is.
But 40W a large difference do not make anyway.

Ok. I'm curious because I'm eyeing the ASUS Prime model for the XT. I like the 2.5 slot design, and it looks neat, but I wouldn't want it beyond ~300W to keep its temps in ideal limits.

 

[adroc_thurston]

but I wouldn't want it beyond ~300W to keep its temps in ideal limits.

It's unlikely to go beyond 300W in general since it's only 64CU.

 

[JustViewing]

AMD should go one up on NVidia by adding custom fixed function hardware to do 8x frame-gen . This will do similar damage to NVidia as Intel's QuickSync did to AMD's hopes for APUs for video compression.

 

[adroc_thurston]

by adding custom fixed function hardware to do 8x frame-gen

There ain't any, DLSS4 MFG is pure ML slop.