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

 

[marees]

more sales for Nvidia

Where are the blackwell cards ???

 

[Timorous]

Where are the blackwell cards ???

At TSMC?

 

[Kronos1996]

Using 1440p TPU data RTX 5080 is 9% better performance per area. While using, in theory, 20W more power and 50% more memory bandwidth.

Hardware Unboxed found it significantly worse than Blackwell. They measure while gaming, not doing synthetics. It’s barely improved over RDNA 3. At least for the 9070 XT.

 

[adroc_thurston]

Hardware Unboxed found it significantly worse than Blackwell. They measure while gaming, not doing synthetics. It’s barely improved over RDNA 3. At least for the 9070 XT.

TPU measures on 3 games too.
Their numbers are an outlier anyway.

 

[SolidQ]

1 card for whole country and premium price

Hardware Unboxed found it significantly worse than Blackwell.

Results so diffrence from game to game

 

[ToTTenTranz]

It runs out of bandwidth all the same.

Only 64MB of Infinity Cache. The chip was made for 1080p to 1440p, never 4K.

GDDR7 on N48 would've been nice for 4K with RT.
Or a bigger chip with more IF and/or more memory channels, of course.

 

[marees]

At TSMC?

I mean, when will they move to the warehouses & retailers ?

If Nvidia goes MIA then AMD is on schedule to hike prices next month

 

[coercitiv]

Only 64MB of Infinity Cache. The chip was made for 1080p to 1440p, never 4K.

Watch the HUB video, there are several instances where the 9070 XT performs better at 4K than it does at 1440p, relative to cards like 7900XTX / 5070 Ti etc.

Common knowledge says the scaling should be worse, and yet the numbers tell a different story (at least partially).

 

[gdansk]

Hardware Unboxed found it significantly worse than Blackwell. They measure while gaming, not doing synthetics. It’s barely improved over RDNA 3. At least for the 9070 XT.

TPU data is only using games too, not synthetics. You can really tell who only looked at HWUB review instead of reading many reviews. From the HWUB review it looks pretty bad. Many others are more positive overall.

I don't know the cause of the discrepancy, perhaps they underclocked again. But as usual expect them to retest it shortly.

 

[Josh128]

I’m thoroughly whelmed. Navi 48 has much worse PPA vs GB203. It’s a 5% small die but 20-25% slower. All I can conclude is that they lost a lot of PPA efficiency by just doubling Navi 44. It should be in-between GB203 + GB205 for this performance level. Closer to 300-330mm2. Also screwed power efficiency. It’s an impressive achievement vs Navi 31/32 but second place is still second place.

Its not terrible, but as we spoke of after the HWUB power leaks, theres still a good bit of distance between it and Blackwell / Ada. However, the chart above is pretty interesting, as is the perf/watt it can achieve from undervolting / power limiting, as shown in the link below.

Radeon RX 9070 XT und 9070 im Test: AMD sprengt die Erwartungen!

Leistungsaufnahme, Effizienz & Tuning

www.pcgameshardware.de

 

[Timorous]

Only 64MB of Infinity Cache. The chip was made for 1080p to 1440p, never 4K.

GDDR7 on N48 would've been nice for 4K with RT.
Or a bigger chip with more IF and/or more memory channels, of course.

If only they made a 96CU model in a 500mm area with a 384 bit bus.

Would likely sit between 4090 and 5090 in perf and could be sold for $1,000. Would have made the 5080 look stupid.

 

[Kronos1996]

TPU measures on 3 games too.
Their numbers are an outlier anyway.

HUB avoids using built-in benchmarks because they’re rarely a good representation of the game. They specifically choose the heaviest parts of the game to use for benchmark runs. Worst case scenario basically but IMO that’s good test methodology. Canned benchmarks tend to be easily manipulated or optimized for marketing purposes. Cough Apple Cough

 

[adroc_thurston]

Only 64MB of Infinity Cache. The chip was made for 1080p to 1440p, never 4K.

RDNA4 introduces a whole bunch of new MALL policies tho.

HUB avoids using built-in benchmarks because they’re rarely a good representation of the game. They specifically choose the heaviest parts of the game to use for benchmark runs. Worst case scenario basically but IMO that’s good test methodology. Canned benchmarks tend to be easily manipulated or optimized for marketing purposes. Cough Apple Cough

a canned game benchmark is just a repeatable slice of the game.
Manual stuff is only ever worth it for, say, MMO CPU benchmarking.

 

[Timorous]

RDNA4 introduces a whole bunch of new MALL policies tho.

a canned game benchmark is just a repeatable slice of the game.
Manual stuff is only ever worth it for, say, MMO CPU benchmarking.

This is where digital foundry have done a good job with their automation. It is essentially their own canned benchmark in custom scenes. Best of both worlds imo.

 

[Kronos1996]

RDNA4 introduces a whole bunch of new MALL policies tho.

a canned game benchmark is just a repeatable slice of the game.
Manual stuff is only ever worth it for, say, MMO CPU benchmarking.

Right and I’m sure no game studios have ever had a reason to create a canned benchmark that makes the game look like it runs way better than it actually does. Surely there would be absolutely ZERO benefit except maybe enticing gamers with older hardware to buy their game expecting a good experience? But they would NEVER do that because game studios are the absolute pinnacle of moral excellence right?

I need to stop before I get a lethal dose of sarcasm. 😂

 

[gdansk]

Hardware Unboxed found it significantly worse than Blackwell.

They also measure it using a proxy, system power in only a few games which you described as heaviest sections of these games. There are sites that use pcat average across all games in their suite and it ends up much closer (to the 5070 Ti). Make of that what you will.

I don't think the 9070 XT is in the efficiency zone for N48 anyway. But it wouldn't be fair to push GB203 down below 300W for a fair comparison because it uses a lot more memory power.

If you go by HWUB alone then even the 9070 isn't efficient despite nearly every other review saying otherwise. Their method might be right. Or it may not be. But look around.

 

[SolidQ]

LM Studio

good comparision between models, length, output etc

9070 XT

docs.google.com

 

[Keller_TT]

The 9070 XT has been pushed beyond its intended scope for that last 3-4% to appear closer to 5070 Ti (and $50↑ price revision), and makes it look dull imo.

The 9070 Pulse and Hellhound are MSRP, near reference designs, and both are very similar in power draw. The Hellhound is cool and quiet too. They come in at 225W sustained, while the 9070 XT is 310W for the Pulse.
The perf gap is a median +13%, and the OC models net a further 2-3% I suppose, but gets into 7900 XTX power levels for worser results in raster.

This smells like Zen 4 frequency maxing. 65W 7700 vs 105W 7700X (97% v 100%). I would take a 97% 9070 XT for 265W.

Edit: Redacted my critical judgement of the +42% perf claims. Not nearly as bad as RDNA3 claims.

 

[itsmydamnation]

LM Studio

good comparision between models, length, output etc

9070 XT

docs.google.com

if your a gamer , thank god it only has 16gb.


Also ROFL at the people on the PPW/PPA anti RDNA4 crusade ..... like so dumb..... it has 40% less bandwidth, that's so much extra performance , normalise for that and look magic now the uarch is wonderful..... like just sigh.....

RDNA4 PPW/PPA is good , sku configs will vary wildly , but lets not pretend that people actually care about PPW.............

 

[SolidQ]

But AMD have outright lied yet again, about +42% vs 7900 GRE.

because it's was mixed RT + Raster

 

[moonbogg]

Where are the blackwell cards ???

If Nvidia decides to waste billions of dollars of shareholder value by allocating wafers for gaming card supply, then people will choose the Blackwell cards over the Radeon cards if both sell at msrp.

 

[adroc_thurston]

If Nvidia decides to waste billions of dollars of shareholder value by allocating wafers for gaming card supply, then people will choose the Blackwell cards over the Radeon cards if both sell at msrp.

DC parts are never wafer-limited.

 

[Keller_TT]

because it's was mixed RT + Raster

It doesn't quite match up, including that. Also see their individual game benchmark numbers. They are beyond margin of error.

 

[gdansk]

If you want a 275W 9070 XT you can probably get that with a slider in the drivers.
Usually neither AMD or Nvidia stop you from setting lower power limits.
And the sites which have tested the 9070 XT at a fixed frame rate as a proxy for equal work efficiency (TPU, CB) usually have it near the top of the charts.

 

[gdansk]

It doesn't quite match up, including that. Also see their individual game benchmark numbers. They are beyond margin of error.

To which numbers are you comparing?
I have checked a few sites and for the same games it is pretty similar...