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]

It's 198mm^2 but the point yeah stands.

Sounds right for 29.7B xtors.

 

[gdansk]

Its like watching the Soviets beat the Americans to space on malnutrition, threats of violence, and watery borscht.

Spoiler

But that analogy is a bit backward. R-7 rocket family was more like Ampere. A very large, brute-force approach compared to smaller, early competitors.

 

[adroc_thurston]

A very large, brute-force approach compared to smaller, early competitors.

that's also how AMD builds GPUs.
They're very straighforward GRUG SMASH devices.

Sounds right for 29.7B xtors.

ayup.

 

[marees]

View attachment 124130

-Another massacred graph but performance looks like it's there. Basically puts the card right in 4070/6800xt territory which ain't shabby.

I figure 16gb will float up to a $400 price on account of the memory, but 8gb will likely hold the line at $300.

Even more interesting if we have a non-xt variant come in at $250 or less down the line. Could literally be a full chip 8gb card that is just clock/power limited to perform around 5060 levels.

Also, how do I get a job in AMD's marketing department... Looks like I could sleep all day and still get a bonus.

$349 for 16GB doesn't sound bad, assuming the performance is still good overall.

If it can beat the 6800 & 7700xt in RT & match the 6800 in raster then that is not a bad deal — at msrp

 

[adroc_thurston]

If it can beat the 6800 & 7700xt in RT

gonna smash 'em.

match the 6800 in raster then that is not a bad deal — at msrp

close nuff

 

[PJVol]

Oh, BTW I am intentionally running this 9070XT on a RM650x (650W Gold) PSU.

At first I was puzzled for a while over the proper connection of this to my 760W Ion+ with just two PCIe cables.

 

[Thunder 57]

It's a separate die with a separate tapeout and a die size of 198mm^2.

I never said it was the same die with parts disabled. That would be a huge waste of area. I see the confusion now. I should have said something like new layout for a 96 bit memory width whereas the smaller die 9060 is basically a halved 9070.

Spoiler

But that analogy is a bit backward. R-7 rocket family was more like Ampere. A very large, brute-force approach compared to smaller, early competitors.
View attachment 124131

The Soviets certainly had better rockets early on and the Atlas wasn't exactly super reliable hence the use of the Redstone for the first two Mercury missions. That all changed when the Saturns came along though. Only recently has the payload capacity of the Saturn V been beaten. And the Soviet attempt at a lunar capable rocket, the N1, well, lol that ended poorly. Luckily they were unmanned.

 

[adroc_thurston]

I should have said something like new layout for a 96 bit memory width whereas the smaller die 9060 is basically a halved 9070.

ah you don't need to, they can just kick 32b mem segment and 4CUs off the N44.

 

[Thunder 57]

ah you don't need to, they can just kick 32b mem segment and 4CUs off the N44.

I suppose so, but I have a feeling that wouldn't be a good idea. The 9060 is probably bandwidth starved as is with 128 bit GDDR6 memory.

 

[adroc_thurston]

The 9060 is probably bandwidth starved as is with 128 bit GDDR6 memory.

Well it's on top of 7700XT and that one is 54CU and 192b GDDR6.
There's leeway for less perf and less money basically.

 

[gdansk]

The Soviets certainly had better rockets early on

That's maybe the point. Ampere is better than RDNA2, after all. But both Ampere & R-7 family are a classic case of using more to do more. Unlike RDNA4 which is somewhat more clever than RDNA2 (and Ampere) in re-ordering and avoiding stalls.

 

[Thunder 57]

Well it's on top of 7700XT and that one is 54CU and 192b GDDR6.
There's leeway for less perf and less money basically.

RDNA4 seems quite a bit more efficient than RDNA3 which is a bit surprising and good to see. As for memory bandwidth that seems to depend on the game. Nvidia benefits greatly from GDDR7 in a few games (Cyperpunk comes to mind) but for others it didn't matter much at all.

Cyperpunk loves memory bandwidth. A more accurate comparison shows a few more games that seem to benefit a bit but most are just showing the "generational" upgrade:

 

[adroc_thurston]

RDNA4 seems quite a bit more efficient than RDNA3 which is a bit surprising and good to see

It's more a case of RDNA3 going horribly-horribly wrong.

Cyperpunk loves memory bandwidth.

Expected, it's a somewhat antique engine from the early deferred era.

 

[SolidQ]

Radeon 9700 is revived

https://videocardz.com/newz/amd-introduces-radeon-ai-pro-r9700-with-32gb-vram-and-navi-48-gpu

 

[Thunder 57]

That's maybe the point. Ampere is better than RDNA2, after all. But both Ampere & R-7 family are a classic case of using more to do more. Unlike RDNA4 which is somewhat more clever than RDNA2 (and Ampere) in re-ordering and avoiding stalls.

I'm not sure I follow. RDNA4 is certainly more clever than RDNA2. The comparison I think you are trying to make is the Soviets used that payload capacity to "brute force" heavier payloads whereas the US made use of lighter payloads (Mercury/Gemini) because they basically had to.

It's more a case of RDNA3 going horribly-horribly wrong.

Expected, it's a somewhat antique engine from the early deferred era.

RDNA3 went very wrong. Zen 2 pulled off chiplets just fine. RDNA3, not so much.

 

[Thunder 57]

Radeon 9700 is revived

https://videocardz.com/newz/amd-introduces-radeon-ai-pro-r9700-with-32gb-vram-and-navi-48-gpu

I had a 9800 Pro and it certainly didn't have any fancy "AI". It did destroy the Geforce FX though.

 

[adroc_thurston]

Zen 2 pulled off chiplets just fine. RDNA3, not so much.

Oh no, chiplets are the only part of RDNA3 that actually worked flawlessly.

 

[SolidQ]

It did destroy the Geforce FX though.

FX 5800 and 9700 fight is back, but in AI now 😁

 

[Thunder 57]

Oh no, chiplets are the only part of RDNA3 that actually worked flawlessly.

What went wrong with it then?

FX 5800 and 9700 fight is back, but in AI now 😁

Didn't even notice that! Is it 2003 again?

 

[adroc_thurston]

What went wrong with it then?

The actual gfx11 microarch and physdes.
They just couldn't get anywhere near targeted clocks.

 

[gdansk]

The comparison I think you are trying to make is the Soviets used that payload capacity to "brute force" heavier payloads whereas the US made use of lighter payloads (Mercury/Gemini) because they basically had to.

Spoiler

I'm not sure. R-7 was a brute-force ICBM. Massive, lots of fuel, surface-launched, easily located, quickly retired in the role for which it was developed. The US built missiles that were just ICBMs. They didn't build a heavy launch vehicle. RDNA4 has no halo part, it's not as much of an ML accelerator. The Soviets weren't doing more with less rocket. They just built a big chungus which is definitely not in line with RDNA4.

 

[adroc_thurston]

They just built a big chungus which is definitely not in line with RDNA4.

well RDNA4 had *the* biggest client chungus imaginable, with composable I/O n stuff.

 

[gdansk]

well RDNA4 had *the* biggest client chungus imaginable, with composable I/O n stuff.

The key part is built. There's not much point to build something you won't sell.

 

[adroc_thurston]

There's not much point to build something you won't sell.

Yuzhmash is littered with stuff like that. lmao.

 

[Thunder 57]

I'm not sure. R-7 was a brute-force ICBM. Massive, lots of fuel, surface-launched, easily located, quickly retired in the role for which it was developed. The US built missiles that were just ICBMs. They didn't build a heavy launch vehicle. RDNA4 has no halo part, it's not as much of an ML accelerator. The Soviets weren't doing more with less rocket. They just built a big chungus which is definitely not in line with RDNA4.

They had powerful rockets early on, but never a "big chungus". They tried with the N1 but it blew up every time. I was wrong about the Saturn V, it still holds the record payload. I thought Starship had surpassed it. SLS and Starship should both in time though. If they had put two or four SRB's on the Saturn V I can't imagine how much more it could've done. But solid fuel wasn't man rated back then for good reason. Look at what happened to two Space Shuttles.

Way off topic though. I bet AMD wish they had made a halo part. I don't think they anticipated Nvidia to crap the bed like they did. And now GN Steve is going after them. He'll never get a review sample again. Might not be a problem if Nvidia insists on "preview" only with many restrictions that no reputable site bothered with. Even Tom's said no. Well, not Tom's Guides.