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

 

[igor_kavinski]

The hardware.

Is RT gonna be 50% faster than RDNA3?

 

[adroc_thurston]

Is RT gonna be 50% faster than RDNA3?

Numbers don't matter, microarchitectural features do.

 

[Kepler_L2]

Is RT gonna be 50% faster than RDNA3?

Depends entirely on the benchmark.

 

[soresu]

Depends entirely on the benchmark.

How about Blender Cycles HIP-RT?

 

[soresu]

Numbers don't matter, microarchitectural features do.

The entire hardware enthusiast community entered the chat.... 😂🤣😆

 

[igor_kavinski]

Depends entirely on the benchmark.

CP2077 and Alan Wake 2?

 

[TESKATLIPOKA]

Is RT gonna be 50% faster than RDNA3?

Numbers don't matter, microarchitectural features do.

I wouldn't care that RT is not any faster than RDNA3, If RDNA4 can clock 50% higher.
The end result would still be the 50% faster RT, but also 50% faster Raster.

 

[SolidQ]

I wouldn't care that RT is not any faster than RDNA3

I think weakest part of AMD RT is Heavy RT and Path Tracing. Light RT is fine.
If RDNA4 like 3-3.5x times faster in Path Tracing, that gonna be on par with 4070ti or slightly faster.

Depends entirely on the benchmark.

If RT like 35% avg, that fine, only question is Path Tracing

 

[Tigerick]

Numbers don't matter, microarchitectural features do.

Since you said about architectural changes, can you confirm the upcoming RDNA4 GPUs going to become RX8000 series? Like RX8600 and RX8700?

 

[SteinFG]

Since you said about architectural changes, can you confirm the upcoming RDNA4 GPUs going to become RX8000 series? Like RX8600 and RX8700?

There's zero chance they're naming a new architecture with the same 7000 series.

 

[soresu]

The end result would still be the 50% faster RT, but also 50% faster Raster.

RT is more memory constrained vs raster.

I don't think simply increasing clocks 50% would be a guaranteed 50% win for RT perf without memory improvements.

 

[Tigerick]

There's zero chance they're naming a new architecture with the same 7000 series.

Except there are lesser and lesser evidence that RDNA4 GPUs are using next gen components:

1. N3E - Initially RDNA4s are rumored to use N3E, later confirmed to use N4P with integrated IF. And there are pretty smaller die size, around 250mm2 for N43.
2. GDDR7 - No longer true cause HVM of GDDR7 only started end of next year and RDNA4s are supposedly debut at early 2024.





I believe most of the leaks are from same source as shown with slide from RGT and also MLID (although he did not mention N43/N44/N48 codenames). I can't help but think it is same hypes as last time about Phoenix was having 9TF of FP32 which ended up half of that.

Without GDDR7, I expect 7700 to come with GDDR6 and 48MB IC, then we could deduct 3 x 37mm2 = 111mm2 from rumored 250mm2. End up about 130mm2 versus 200mm2 of N32 (60 CU/RA), we could guess how many CU new die of RDNA4 is having...

 

[Tigerick]

CP2077 and Alan Wake 2?

How about I show you the potential CP2077's RT performance figures instead of the pep-talk others are giving:-

Here's TPU performance review of Phantom Liberty, I summarized the figures below:-

1920 x 1080 RT	RT cores	FPS	FPS / RA
RX 7900XTX	2nd Gen 96CU	49.7	0.52
RX 7900XT	2nd Gen 84CU	44.1	0.525
RX 7800XT	2nd Gen 60CU	34.6	0.58
RX 7700XT	2nd Gen 54CU	30.9	0.57
RX 7700 ?	3rd Gen 40CU	~ 25.6	0.58 * 1.1 = 0.64
RX 6700XT	1st Gen 40CU	21.1	0.53
RX 7600XT ?	3rd Gen 32CU	~ 20.5	0.58 * 1.1 = 0.64
RX 7600	2nd Gen 32CU	14.8	0.46
RX 6600XT	1st Gen 32CU	13.4	0.42

• Even though RX7600 comes with 2nd Gen of RT cores, the performance is similar to RX6600; they are far slower than 6700XT with 1st Gen of RT cores. I think @TESKATLIPOKA has questioned it before, that's why there is no relationship between generation of RT cores...the same could be said to Shader ISA, it is matter of product positioning. AMD won't make 7600's RT performance faster than 6700XT cause AMD still expect to sell them for a while. Therefore, all the talks about Shader ISA, micro architectural changes are non-sense compared to product positioning, remember that.
• AMD won't ditch 7700XT and 7800XT which are just released half a year ago, thus N43 and N44 are positioned below N32 not above. Even though N43/44 are using newer Shader ISA and 3rd gen of RT cores, I expect 3rd gen of RT cores are having similar performance of N32's 2nd Gen with maybe 10% better performance due to higher clock speed.
• RX7600XT with 32CU/RA cores will most likely perform 38% better compared to RX7600, hence the XT naming. The performance has almost reached to 6700XT. Here is the source of better RT performance, it is all due to product positioning...
• OTOH, RX7700 with 40CU/RA cores will perform better than 6700XT but slower than 7700XT. AMD won't jeopardize sales of N32 cards, thus 40CU seems reasonable amount of CU instead of 48CU. That's why don't expect much performance improvement of RX7700 compared to RX7700XT.

 

[Tigerick]

I remember @DisEnchantment has questioned about the positioning of the RDNA4 before:-

I am just wondering why replace lower end/midrange parts with new architecture? I would expect current lower midrange 7700XT next year will be coming down to 399USD levels. Once NV adjust 4060 prices it will come down.
Which means they don't need RDNA4 at all for next year. They could just create a Strix successor and put RDNA4 inside for 16CUs and call it a day with RDNA4.

And I think I have answers for him, if you guys still not convince, well more signs will follow in the following month...

 

[SteinFG]

Except there are lesser and lesser evidence that RDNA4 GPUs are using next gen components:

1. N3E - Initially RDNA4s are rumored to use N3E, later confirmed to use N4P with integrated IF. And there are pretty smaller die size, around 250mm2 for N43.
2. GDDR7 - No longer true cause HVM of GDDR7 only started end of next year and RDNA4s are supposedly debut at early 2024.


View attachment 89399


I believe most of the leaks are from same source as shown with slide from RGT and also MLID (although he did not mention N43/N44/N48 codenames). I can't help but think it is same hypes as last time about Phoenix was having 9TF of FP32 which ended up half of that.

Without GDDR7, I expect 7700 to come with GDDR6 and 48MB IF, then we could deduct 3 x 37mm2 = 111mm2 from rumored 250mm2. End up about 130mm2 versus 200mm2 of N32 (60 CU/RA), we could guess how many CU new die of RDNA4 is having...

It's not about the node or the number or CUs or mm of silicon, but the architecture

 

[adroc_thurston]

Except there are lesser and lesser evidence that RDNA4 GPUs are using next gen components:

1. N3E - Initially RDNA4s are rumored to use N3E, later confirmed to use N4P with integrated IF. And there are pretty smaller die size, around 250mm2 for N43.
2. GDDR7 - No longer true cause HVM of GDDR7 only started end of next year and RDNA4s are supposedly debut at early 2024.

You're saying wordswords yet they don't mean anything.

It's not about the node or the number or CUs or mm of silicon, but the architecture

Bingo!

 

[Tigerick]

You're saying wordswords yet they don't mean anything.

Bingo!

Since you said about architectural changes, can you confirm the upcoming RDNA4 GPUs going to become RX8000 series? Like RX8600 and RX8700?

 

[SteinFG]

Since you said about architectural changes, can you confirm the upcoming RDNA4 GPUs going to become RX8000 series? Like RX8600 and RX8700?

Deja Vu, I've just been in this place before.
Higher on the street, And I know it's my time to go.
Calling you, and the search is a mystery.
Standing on my feet, It's so hard when I try to be.

 

[Ajay]

It's not about the node or the number or CUs or mm of silicon, but the architecture

They all matter (although #CUs and mm2 are related). The crux is that a bad architecture cannot be fixed. A poor implementation may be able to be fixed - it's a cost/benefit problem. Aside from a situation where a respin of one layer has large gains in yields, it's generally not worth it with today's very expensive validation process and tighter TTM targets.

 

[SteinFG]

They all matter (although #CUs and mm2 are related). The crux is that a bad architecture cannot be fixed. A poor implementation may be able to be fixed - it's a cost/benefit problem. Aside from a situation where a respin of one layer has large gains in yields, it's generally not worth it with today's very expensive validation process and tighter TTM targets.

I am talking about a completely different topic

 

[adroc_thurston]

Since you said about architectural changes, can you confirm the upcoming RDNA4 GPUs going to become RX8000 series? Like RX8600 and RX8700?

Numbers don't matter.

 

[Tigerick]

Numbers don't matter.

Ha, so you have been BSing all the times now said numbers don't matter. Thanks man, I think I know the answers already. If others can't read between the lines, well the launching of RDNA4 cards will tell the full story

 

[adroc_thurston]

Ha, so you have been BSing all the times now said numbers don't matter.

You're trying to attach final numbers to a piece of Si that hasn't PRQ'd.
You're being silly (more dumb than silly but I digress).

 

[Tigerick]

You're trying to attach final numbers to a piece of Si that hasn't PRQ'd.
You're being silly (more dumb than silly but I digress).

I am just asking for model numbering, be it RX7000 or RX8000? Is that so hard for you to answer? Or simply I don't know also matter...

 

[SteinFG]

I am just asking for model numbering, be it RX7000 or RX8000? Is that so hard for you to answer? Or simply I don't know also matter...

It's definitely not 7000 💀