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

 

[SolidQ]

AMD missing this huge opportunity by canning the big RDNA4 is nuts

at least they should make ~4090 level for 1k, but again miss oportunity

 

[sandorski]

AMD missing this huge opportunity by canning the big RDNA4 is nuts. RDNA4 looks like an actually successful, competitive, good arch with great engineering and they just squandered this huge opportunity. Who knows if RDNA5 will have the same chance vs Rubin or whatever the next NV arch is.

They can own the Bulk of the Market. Win mindshare as much as Marketshare. Over priced Halo cards can come another day and they will.

 

[eek2121]

OoO is by itself far more complicated to implement performantly and efficiently than in order compute.

Especially given the number of engineers that have worked on OoO GPU designs is far less than those which have worked on OoO CPUs in the history of the industry.

You're just substituting process/packaging complexity for µArch complexity - which kinda runs counter to the whole chiplet era ethos of keeping die sizes down for minimised mask costs and maximised yields, something they had already identified as an issue before Zen 2 was in development 7-9 years ago.

Gotta disagree with this. There are some workloads that will benefit massively, and the rest of what you are describing is developers working around hardware limitations. When If OoO lands in consoles, that WILL change.

So we're not there yet. Performance looks similar to a 4080. I saw some videos on youtube and it's enough for RT medium aka normal RT. For path tracing, even 1080p you need a 5090.

Depends on the game and the settings.

 

[DAPUNISHER]

30min please don't ban me im, a good boy
View attachment 118843

I hearted this post when you made it on X.

 

[eek2121]

AMD missing this huge opportunity by canning the big RDNA4 is nuts. RDNA4 looks like an actually successful, competitive, good arch with great engineering and they just squandered this huge opportunity. Who knows if RDNA5 will have the same chance vs Rubin or whatever the next NV arch is.

Meh, most people don’t pay over $600 for GPUs.

I would like to see a halo product (it has been a long time since my gaming system had an AMD GPU), however, they will get there, eventually.

Funny thought I had writing this: I own more AMD GPUs than NVIDIA ones, and that doesn’t include APUs.

EDIT: and after NVIDIA's black screen nonsense (which I suffered severely from), I would definitely purchase an AMD upgrade to my 4090, when one exists.

 

[soresu]

Gotta disagree with this. There are some workloads that will benefit massively, and the rest of what you are describing is developers working around hardware limitations. When If OoO lands in consoles, that WILL change.

My point was about the actual engineering experience in the industry related to OoO GPUs, and the general difference in µArch design/implementation complexity vs InO at the hardware level.

Comparatively I guess the same can be said about chiplet GPU architectures focused on gaming.

 

[soresu]

30min please don't ban me im, a good boy
View attachment 118843

Replace Loki's face with Jensen's and you've got a deal 😆

 

[soresu]

Funny thought I had writing this: I own more AMD GPUs than NVIDIA ones, and that doesn’t include APUs.

Strix Halo is looking very good to me for a future SFF PC build.

I like not so big PC butts and I cannot lie 😆

Though the fact that it lacks RDNA4's RT and VCN upgrades is enough to have me wait until Medusa Halo to buy a hi end APU system.

 

[DAPUNISHER]

Replace Loki's face with Jensen's and you've got a deal 😆

That would be cool, but I'm already stoked enough you guys are becoming meme lords, just like me. If we can't have fun with it, what's the point?

If anyone with a 50 series could fire up some old DX9 games and report back -

 

[eek2121]

Strix Halo is looking very good to me for a future SFF PC build.

I like not so big PC butts and I cannot lie 😆

Though the fact that it lacks RDNA4's RT and VCN upgrades is enough to have me wait until Medusa Halo to buy something with it.

A Strix Halo mini PC is on my buy list for later this year.

 

[techjunkie123]

I want a strix halo laptop, 14-16 portable workstation, but there are no good options .

 

[Josh128]

AMD missing this huge opportunity by canning the big RDNA4 is nuts. RDNA4 looks like an actually successful, competitive, good arch with great engineering and they just squandered this huge opportunity. Who knows if RDNA5 will have the same chance vs Rubin or whatever the next NV arch is.

This is all pure speculation until actual 3rd party reviews and tests drop. AMD likely had very good reasons for what they are doing. We just dont know exactly what they are, at this point.

 

[Panino Manino]

RTX 5070 is really bad, and look, Steve slipped the 9070 and 9070XT power consumption in the review.

 

[Panino Manino]

Meh, most people don’t pay over $600 for GPUs.

But some people do.
Produce a "symbolic" quantity of BIG chips for those who will surely pay to get some of the "mindshare" back.

 

[SolidQ]

9600XT is N44 or N48 that is question

https://twitter.com/x/status/1896921459660112307

 

[linkgoron]

This is all pure speculation until actual 3rd party reviews and tests drop. AMD likely had very good reasons for what they are doing. We just dont know exactly what they are, at this point.

Yes you're right of course, I was planning on adding "if AMD's benchmarks match 3rd part benchmarks" but forgot. My bet is that AMD just didn't expect Blackwell to be such a misstep/disaster. They thought that Nvidia can do no wrong.

 

[Josh128]

RTX 5070 is really bad, and look, Steve slipped the 9070 and 9070XT power consumption in the review.


View attachment 118849
View attachment 118850

Those power consumption figures for RDNA 4 dont look good in comparison. 9070 draws more power than 5080? Looks like 9070XT will have next to no perf/watt uplift vs 7900XT. Who was saying that AMD matched NV in perf/watt/area?

**Also, why is 5070 drawing more power than 5080 in Warhammer?

 

[gdansk]

Who was saying that AMD matched NV in perf/watt

It will depend on the game. TPU, for example, recently changed their power consumption methodology. Coincidence, possibly.
But if they can't reduce power consumption in typical games it may explain why it doesn't appear in laptops.

 

[SolidQ]

Looks like 9070XT will have next to no perf/watt uplift vs 7900XT. Who was saying that AMD matched NV in perf/watt/area?

Very different results from game to game

 

[branch_suggestion]

Those power consumption figures for RDNA 4 dont look good in comparison. 9070 draws more power than 5080? Looks like 9070XT will have next to no perf/watt uplift vs 7900XT. Who was saying that AMD matched NV in perf/watt/area?

**Also, why is 5070 drawing more power than 5080 in Warhammer?

GPU power consumption at stock can vary wildly depending on what is being hammered and if it is fully stressed on the shaders or not.
Furmark is king for good reason.

 

[MrTeal]

Those power consumption figures for RDNA 4 dont look good in comparison. 9070 draws more power than 5080? Looks like 9070XT will have next to no perf/watt uplift vs 7900XT. Who was saying that AMD matched NV in perf/watt/area?

**Also, why is 5070 drawing more power than 5080 in Warhammer?

Yeah, that is a massive increase in power over the Blackwell cards.

Hopefully some of it is due to an increase in EPS draw from better CPU utilization.

 

[eek2121]

They are measuring total system power maybe? Either that or an OC card? The TGP is only 304W.

Another minor annoyance with most reviewers is they don’t mention the monitor characteristics or how many monitors are attached to the system.

Power consumption will differ based on refresh rate, monitor resolution (which may be different from game resolution), etc.

 

[eek2121]

Who was saying that AMD matched NV in perf/watt/area?

I did, but that is based on TGP and what little info we know.

I’d like to see measurement of the GPU itself isolated while running something like furmark.

 

[gaav87]

I did, but that is based on TGP and what little info we know.

I’d like to see measurement of the GPU itself isolated while running something like furmark.

I showed furmark results... I guess we have another rdna2 ? Soft shows 330w wall 400w xD

 

[GTracing]

Those power consumption figures for RDNA 4 dont look good in comparison. 9070 draws more power than 5080? Looks like 9070XT will have next to no perf/watt uplift vs 7900XT. Who was saying that AMD matched NV in perf/watt/area?

**Also, why is 5070 drawing more power than 5080 in Warhammer?

I said it. Specifically that RDNA4 is matching Blackwell on an architectural basis. Based off HWU's numbers it looks like RDNA4 is ~10% behind in PPA.

To summarize the differences between N48 and GB203

• GB203 uses 22mm^2 more die area. 378mm^2 VS 356mm^2. It's 6% bigger.
• GB203 uses less power. The 5080 is ~40-50 watts less than the 9070 XT. Or roughly 15% less
• GB203 offers more gaming performance. The 5080 is probably ~15-20% higher than the 9070 XT. We'll know for sure tomorrow.
• GB203 uses GDDR7 with 40% more bandwidth than the GDDR6 that N48 uses.
• They might use slightly different process nodes. I don't know how this would affect things.

Factoring in the memory difference is tricky, but if we assume that 64MB of cache would provide a similar amount of memory bandwidth, we could say that it's the equivalent of ~40mm^2 of die area.

Power consumption is exponential with area. So normalized to the same wattage the 15% power difference becomes a 5-10% performance difference.

In my opinion RDNA4 is a huge improvement. They're a bit behind Nvidia, but it's the closest they've been in forever.