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

 

[KompuKare]

Somewhat back on topic, it will be interesting to me at least if the PS5 Pro with the shiny new GPU can slot in at or below the power and cooling envelope that the launch PS5 had.

I am making some assumptions, but we are moving from 7nm TSMC to… 5nm TSMC? I know the PS5 made an intermediate move to N6. This along with whatever PPW improvements the architecture includes is what has me curious, as the CPU block seems mostly unchanged and the process efficiency improvements seem to be feeding a larger GPU power & thermal budget. Maybe.

Weren't the rumours during the shortages and PS5 yields were poor - and further rumours that Sony had upped clocks late in the deployment cycle (probably after finding out that XBX was a bigger die) and this was a major reason for the poor yields.
I'm sure AMD learned some lessons so maybe despite being on a newer and more expensive node, they could have made it slightly wider - thereby solving they power and yield problems?

 

[blckgrffn]

Weren't the rumours during the shortages and PS5 yields were poor - and further rumours that Sony had upped clocks late in the deployment cycle (probably after finding out that XBX was a bigger die) and this was a major reason for the poor yields.
I'm sure AMD learned some lessons so maybe despite being on a newer and more expensive node, they could have made it slightly wider - thereby solving they power and yield problems?

I think you are right about what happened at launch, and Sony supposedly said they were good with the lower yields if they got a more competitive chip late in the game. Then Covid + mining, of course. Given how it was a drop in replacement for the PS4 who knows how it might have done without those supply issues. But still, its impressive in its capability and sound and power envelope.

It looks like the new GPU chip is significantly wider and not clocked that much higher, and the N4 node should be very mature, so yields should be good? I guess I am curious if whatever was plaguing RDNA3 from the standpoint that it seemed hungrier than AMD anticipated when clocks ramped was addressed of if it is simply clocked low enough on good silicon that this problem was minimized via different methods.

It looks like a step from a 6700 (non XT) to 7700XT in raster and maybe up to 4070 levels of RT performance in a reasonable power envelope, which is hopefully a good sign for RDNA4 taking that much bigger of step.

The CPU is very similar to Zen 2, has that ever been ported down to sub N6 before?

 

[Rekluse]

Pricewise is RDNA4 expected to be less than or equal to the current gen ? eg

Price
RX8800 <= RX7800
RX8700 <= RX8700
RX8600 <= RX8600

etc

Given the smaller die's and lack of Halo

 

[Kepler_L2]

Pricewise is RDNA4 expected to be less than or equal to the current gen ? eg

Price
RX8800 <= RX7800
RX8700 <= RX8700
RX8600 <= RX8600

etc

Given the smaller die's and lack of Halo

There is no 8800.

 

[jpiniero]

It looks like the new GPU chip is significantly wider and not clocked that much higher

Apparently the 2.35 Ghz max clock only happens in rare circumstances to the point that Sony may not even advertise it as being that high.

 

[PJVol]

There is no 8800.

No point in marketing ploys if expectations are met?

 

[Tuna-Fish]

There is no 8800.

Do you already know the naming of the cards? How is the stack split? I would assume there would be two cuts of each die, and that would work fairly nicely with 800/700 and 600/500.

If the leaked specs are right, I would expect it to always beat a 7800XT, and even if it didn't, AMD has launched a (n+1)800 that was weaker than (n)800 before. (HD 6800)

 

[Kepler_L2]

Do you already know the naming of the cards? How is the stack split? I would assume there would be two cuts of each die, and that would work fairly nicely with 800/700 and 600/500.

If the leaked specs are right, I would expect it to always beat a 7800XT, and even if it didn't, AMD has launched a (n+1)800 that was weaker than (n)800 before. (HD 6800)

Technically name and price can change right before announcement, but AFAIK it will look something like this:

N48:
8700 XTX 16GB
8700 XT 16GB
8600 XT 12GB
N44:
8500 XT 8GB

 

[SolidQ]

N48:
8700 XTX 16GB
8700 XT 16GB
8600 XT 12GB
N44:
8500 XT 8GB

perf like this?
~ 7900XT
~ 7800XT
~7700XT
~4060ti
Why AMD/NV love degradation in Vram(mid/low market)
6800 16gb> 7700XT/8600XT 12gb
3060 12 > 4060 8
7600XT 16 > 8500XT 8

 

[Timorous]

Technically name and price can change right before announcement, but AFAIK it will look something like this:

N48:
8700 XTX 16GB
8700 XT 16GB
8600 XT 12GB
N44:
8500 XT 8GB

Not keen on that naming scheme but whatever. That 8gb 8500XT will get obliterated in reviews if pricing is over $220.

 

[igor_kavinski]

N48:
8700 XTX 16GB
8700 XT 16GB
8600 XT 12GB
N44:
8500 XT 8GB

That's a good enough line-up. Wish they had done that with RDNA3. 7600 8GB should've been 7500 XT and cheaper. 7600 XT 12GB should've been in place of plain 7600 and so on.

 

[igor_kavinski]

Why AMD/NV love degradation in Vram(mid/low market)
6800 16gb> 7700XT/8600XT 12gb
3060 12 > 4060 8
7600XT 16 > 8500XT 8

RX 6800 having no RDNA3 analogue is purely the fault of reviewers for recommending 6800 XT over it.

3060 12GB became a thorn in Nvidia's side when it started beating 3060 Ti in some games so they weren't gonna repeat that "mistake" of giving more for less.

8500 XT 8GB should be fine if it's $175 or less.

 

[marees]

Not keen on that naming scheme but whatever. That 8gb 8500XT will get obliterated in reviews if pricing is over $220.

8500 XT 8GB should be fine if it's $175 or less

If 4060 ti is $400 then how much can AMD price its 8gb RDNA 4 card (that will have matching raster & RT performance) ?

 

[marees]

Do you already know the naming of the cards?

How about Radeon RX AI 170 ?

 

[Timorous]

If 4060 ti is $400 then how much can AMD price its 8gb RDNA 4 card (that will have matching raster & RT performance) ?

It's 8GB in 2024. That gives it a ceiling regardless of performance when not at the VRAM limit because more and more titles have performance or IQ issues with just 8GB.

If performance is 4060Ti like then I think AMD could get better margin just by giving it more VRAM.

 

[Aapje]

Technically name and price can change right before announcement, but AFAIK it will look something like this:

N48:
8700 XTX 16GB
8700 XT 16GB
8600 XT 12GB
N44:
8500 XT 8GB

If the 5070 Ti is ~7900 XT is ~8700 XTX, then they line up with Nvidia again. I can how it is hard for them to call it the 8800 when the 4080 already beats the 7900 XT and 5080 is going to be faster still.

And does AMD think that Nvidia will offer a good price/perf improvement for the 5060? Because it seems like the profit margin for the 8600 XT will be problematic if it is N48.

 

[igor_kavinski]

If 4060 ti is $400 then how much can AMD price its 8gb RDNA 4 card (that will have matching raster & RT performance) ?

I think they may launch it at $349 and then ultimately settle at around $279 once sufficient volume is shipped. But Nvidia responding with a price cut could ruin their margins a fair bit.

 

[marees]

I think they may launch it at $349 and then ultimately settle at around $279 once sufficient volume is shipped. But Nvidia responding with a price cut could ruin their margins a fair bit.

I don't expect a 8gb card to launch for more than $300. That way nvidia also has to climb down by a big amount on the 4060 ti

 

[jpiniero]

IMO, I think AMD should only release one SKU per die, at least to start. I doubt that happens though.

 

[igor_kavinski]

I don't expect a 8gb card to launch for more than $300.

It depends on what RDNA2 cards it may beat. If it beats a 6750 XT in less VRAM intensive games in both raster and RT, AMD could pull a Jensen move and try to justify the higher performance with a higher price tag.

 

[TESKATLIPOKA]

perf like this?
~ 7900XT
~ 7800XT
~7700XT
~4060ti
Why AMD/NV love degradation in Vram(mid/low market)
6800 16gb> 7700XT/8600XT 12gb
3060 12 > 4060 8
7600XT 16 > 8500XT 8

Difference between 7900XT vs 7700XT is 52% at 1440p, so that's very very unlikely.

 

[Mopetar]

IMO, I think AMD should only release one SKU per die, at least to start. I doubt that happens though.

Maybe for N44 that's possible since it's so small that most dies will be good and the binned dies can be sold to OEMs who want a cheap, low power discreet card for some PCs.

For N48 they're going to be at least two parts just to handle any parts with defects or lower performance characteristics.

 

[ToTTenTranz]

8500 XT 8GB

This had better be less than $170.
8GB are inexcusable nowadays.

 

[adroc_thurston]

This had better be less than $170.

Man this is pretty raw optimism.

8GB are inexcusable nowadays.

DRAM stopped scaling.
Excuses or not, it's what you get.

 

[jpiniero]

DRAM stopped scaling.
Excuses or not, it's what you get.

More than 2 GB GDDR chips is on the roadmap... just not yet.