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

 

[branch_suggestion]

I thought the PS5 used GDDR6? Is the Pro's RAM clocked too low?

It is 18Gbps. Fine for what it is, but at desktop speeds you would need more bandwidth from somewhere.

 

[marees]

I thought the PS5 used GDDR6? Is the Pro's RAM clocked too low?

18mbs to allow competitive bidding across different memory manufacturers

Must be same reason AMD prefers it for its mainstream GPUs

 

[marees]

Sony add AI upscaling on AMD, hope AMD going with RDNA4 also for AI upscaling, otherwise it's gonna be strange, Sony can do it on AMD GPU, AMD can't

AMD Debuts New SoC With 60 Compute Units, Faster GDDR6 Memory, Advanced Raytracing & PSSR AI Upscaling In PlayStation 5 Pro

AMD developed a new custom SoC for Sony's PlayStation 5 Pro, featuring 60 Compute units based on the RDNA architecture & other improvements.

wccftech.com

Another major feature being announced today is PSSR or PlayStation Spectral Super Resolution which is an AI-driven upscaling method. XDNA 2, the same IP that is powering the NPU for AMD's Strix Point APUs will be used to handle the AI processes on the PS5 Pro.

Wait for Paul Alcorn's articles. He is not finished yet.

 

[adroc_thurston]

hope AMD going with RDNA4 also for AI upscaling

Nope.

Wish they had souped it up with 64MB MALL cache. Would've made the price worth it.

too much money

 

[Hans Gruber]

18mbs to allow competitive bidding across different memory manufacturers

Must be same reason AMD prefers it for its mainstream GPUs

The problem is the 4070 uses 21gbps GDDR6X memory. The other problem is there is a new GDDR6 non X variant that uses 20GBit memory on a new variant of the 4070. That GPU is 2 years old and AMD is going to use 18GBit memory in 2024/2025? It makes no sense to me. Why doesn't AMD follow the new 4070 variant with 20Gbit GDDR6 memory?

 

[Mahboi]

AMD is a poor company unlike rich Sony that can sell consoles in the range of $800-$1000

I laughed way more than I should have, thanks

 

[adroc_thurston]

Why doesn't AMD follow the new 4070 variant with 20Gbit GDDR6 memory?

cost

 

[marees]

The problem is the 4070 uses 21gbps GDDR6X memory. The other problem is there is a new GDDR6 non X variant that uses 20GBit memory on a new variant of the 4070. That GPU is 2 years old and AMD is going to use 18GBit memory in 2024/2025? It makes no sense to me. Why doesn't AMD follow the new 4070 variant with 20Gbit GDDR6 memory?

As I said, 18gbps is the lowest common denominator across memory companies. So AMD is bidding for PS5, PS5 pro, RDNA 3 & RDNA 4. They probably get good discounts.

Hopefully once the RDNA 3 stocks clear, you might get cheap RDNA 4 cards.

Imagine if AMD's 4070 gddr6 competitor costs only $350. Can nvidia match it ?

 

[igor_kavinski]

too much money

Oh come on. Cache is cheap enough that AMD is putting it on 5600X3D and 5700X3D.

 

[SolidQ]

Nope.

Than zero chance to get market share, while doesn't have competitor to DLSS. People gonna buy NV still even if Radeon 50% faster for same price, because DLSS

 

[igor_kavinski]

because DLSS

I hope RDNA4 fixes that.

 

[SolidQ]

I hope RDNA4 fixes that.

Same, the problem FSR is ok for 1440+, but if you looking steam there like 56% 1080p, and and more % gpu's for 1080p. That big problem for AMD, because FSR falls off in 1080p(expect few games) compared to DLSS

 

[adroc_thurston]

Cache is cheap enough that AMD is putting it on 5600X3D and 5700X3D.

These are obscene margins by console standards.

Than zero chance to get market share, while doesn't have competitor to DLSS. People gonna buy NV still even if Radeon 50% faster for same price, because DLSS

They just have to bribe the shills to shill raw uarch PPA advantage.
It takes one check mailed to Battaglia and there you go.

I hope RDNA4 fixes that.

No.
Matrix math is Haram unless you need to hit SOC tops targets mandated by MS.

 

[igor_kavinski]

Matrix math is Haram unless you need to hit SOC tops targets mandated by MS.

I see you got some Jihadi friends fighting alongside you

 

[adroc_thurston]

I see you got some Jihadi friends fighting alongside you

Of you wouldn't believe me.

Sheikh Mansur Battalion - Wikipedia

en.wikipedia.org

Anyway, back to RDNA4.

 

[SolidQ]

Battaglia and there you go.

it's not only him, it's most trusted youtubers now

They just have to bribe the shills to shill raw

NVIDIA already plague market with DLSS, so AMD need same answer., without good competitor to DLSS, not happens, let's alone RT

 

[adroc_thurston]

it's not only him, it's most trusted youtubers now

Patient zero is the first to mail checks for.

NVIDIA already plague market with DLSS, so AMD need same answer., without good competitor to DLSS, not happens, let's alone RT

All it takes is shills singing the praise to raw PPA glory.
This would've happened with RDNA3 but they flubbed. Hard.

 

[Tuna-Fish]

The problem is the 4070 uses 21gbps GDDR6X memory. The other problem is there is a new GDDR6 non X variant that uses 20GBit memory on a new variant of the 4070. That GPU is 2 years old and AMD is going to use 18GBit memory in 2024/2025? It makes no sense to me. Why doesn't AMD follow the new 4070 variant with 20Gbit GDDR6 memory?

Even if PS5 Pro sells badly for a console, it will probably outsell all GPUs with 20Gbps GDDR6 put together in fairly short order. The highest speed grade of a memory type is just generally not usable for such a high-volume product. Because if you had tons of ram hitting 20Gbps speeds, you'd, by the virtue of the bell curve, have a bunch or ram that could do better. And then that would be the highest speed grade. The fact that you can't buy 22Gbps GDDR6 probably means there isn't enough 20Gbps GDDR6 for a console.

 

[gaav87]

PS5 PRO - NOV 7
AMD AI event - OCT10
...
RDNA4 OCT10 technical presentation NOV launch ?

 

[jpiniero]

They just have to bribe the shills to shill raw uarch PPA advantage.

You and I know that nVidia is going to hype the hell out of Blackwell's AI performance. Maybe that's the real reason for going back to one uarch... because they are behind there.

 

[Shmee]

Still beating with new

Spoiler

Wow, that is still impressive. Though I stopped playing much COD a while back, I even uninstalled it. I really liked the MWII groundwar mode, I think that was my peak COD experience recently. But the MW3 only has 3 groundwar maps, not really anything special there, and I am not interested in the coming BlackOPs title.

 

[adroc_thurston]

You and I know that nVidia is going to hype the hell out of Blackwell's AI performance. Maybe that's the real reason for going back to one uarch... because they are behind there.

They ugh don't have area to spam fatter GEMM cores in client.

 

[ToTTenTranz]

Even if PS5 Pro sells badly for a console, it will probably outsell all GPUs with 20Gbps GDDR6 put together in fairly short order.

If the public reception is any indication (and it usually is), at the current prices the PS5 Pro is going to sit on the shelves with horrid reception and Sony's marketing team severely overestimated how much the die hard fans are willing to pay for a console that simply plays games in Quality Mode with Performance Mode framerates.
Not to mention the brand perception crisis they just created.

The introduction of RDNA4 discrete GPUs is going to make everything worse.

 

[jpiniero]

They ugh don't have area to spam fatter GEMM cores in client.

There's a reason GB205 and GB207 are cutting core counts...

The introduction of RDNA4 discrete GPUs is going to make everything worse.

You know people are going to complain about RDNA4's prices. That may end up being a problem now too since people apparently have taken what AMD said that it'll be $100 for the top model.

 

[blackangus]

Than zero chance to get market share, while doesn't have competitor to DLSS. People gonna buy NV still even if Radeon 50% faster for same price, because DLSS

I have a 3070ti and have used both FSR and DLSS. I honestly cant tell the difference at 1440P or 4K. Both have little artifacts, but IMHO it just doesn't appear to be a big difference at all.
IMHO the no DLSS competitor isnt really true.
On a few games I ended up using FSR to get frame gen were my poor 3070 TI isnt even supported by NVidia for the feature.