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

 

[PJVol]

So we ignore the 56 vs 64CUs (+14%)

We're not, but one have to keep in mind that games in general are the worst-case scenario for evaluating arch scaling, and at least it make sense to wait for the drivers to mature.
Btw the difference in performance (acc. to AMD) is 15% at 1440p and 17.3% at 4K. Although the power scales linearly with CU number and gaming clock

All Powercolor models use Honeywell PTM7950

Sapphire too.

 

[Bryo4321]

Is there confirmation of XFX using Honeywell? They did on the XTX mag air…

 

[blckgrffn]

Is there confirmation of XFX using Honeywell? They did on the XTX mag air…

Right? This is great news on the other manufacturers! It was the defining reason I spent extra money on my XTX vs the Pulse, I just wanted to not have to mess with it down the road as I keep my GPUs for a bit.

 

[gaav87]

They lost their mind

 

[Jan Olšan]

Then why are you this thread? Weird flex. Also, good job disregarding a reasonable request not to use rather ugly language when asked. How is that 40% ipc doing?

Shh, but AMD says a single RDNA 4 compute unit is +40% over a single RDNA 3 Compute Unit. Not just IPC I think, it's with clock boost included.

 

[Jan Olšan]

It's still getting the "Editor's Choice" and "Recommended" badges over at TechPowerUp, no matter what.

Speaking from first-hand experience: It's a bit complicated with the awards badges. Of course, companies want them badly, even to points where t makes no sense (paying literally whos for their literally-who-stamp-of-approval). By paying I mean "Nvidia/AMD/AIB: you can keep the GPU after review for your test lab" (and this is probably p. fine) "literally who: off to e-bay you go" Nvidia/AMD/AIB: ignores because badges

As for the expensive products - it's a conundrum whether you should shit on a product that works well, is good quality etc, no big issues, but it's expensive. NOctua NH-D15G2 for example.

So one of the compromise solutions is to give the card whatever functional award from your catalogue you think it deserves, but add another stamp that adds "But this SHIT is too expensive". I think TPU does exactly that.

It basically is supposed to mean that yeah it is nice, but you shouldn't buy it because price always matters. Of course, some people will only look at the nice-stuff badge, ignoring the flipside one.

 

[Timorous]

They lost their mind

View attachment 118388

Have you seen the prices at OCUK. They are £12,000...

 

[gaav87]

Have you seen the prices at OCUK. They are £12,000...

but u can buy them now in poland or nl for 1200euro shipping from amazon nl tmr and poland on 6th to me

 

[SolidQ]

They are £12,000.

12k or 1.2k?

Sapphire Nitro+

 

[moonbogg]

but u can buy them now in poland or nl for 1200euro shipping from amazon nl tmr and poland on 6th to me

Is 1200 euro good or bad.

 

[gaav87]

Is 1200 euro good or bad.

idk 120% over 549$ msrp is good or bad ?

 

[Timorous]

12k or 1.2k?

Sapphire Nitro+

12k. It is clearly a temporary price though which they will update when you can actually buy them.

 

[MrTeal]

but u can buy them now in poland or nl for 1200euro shipping from amazon nl tmr and poland on 6th to me

Amazon always has demand pricing, you can't make any inferences from what they want for an unreleased product a week before launch.

 

[Philste]

We're not, but one have to keep in mind that games in general are the worst-case scenario for evaluating arch scaling, and at least it make sense to wait for the drivers to mature.

While that is generally true, we shouldn't wait for drivers to mature. AMD is a multi billion Dollar company, if drivers are garbage again than that is close to unacceptable. What you wrote basically sounds like we are at a point where it is acceptable for AMDs drivers to be bad at the beginning and it's great when they mature sometime in the future. However that is not what we should expect from a company like that.

Also looking at RDNA2: 6900XT had 11% more CUs than 6800XT without any other changes: Same clocks, same TDP. Yet it delivered like 8-9% perf at 4K over 6800XT at that same TDP. So RDNA4 scaling seems to be quite meh for now, we will see how N44 turns out later this year.

For now it could be a first indication why other RDNA4 was cancelled, at least seems more likely than just "GB202 too strong, we give up".

 

[coercitiv]

Right? This is great news on the other manufacturers! It was the defining reason I spent extra money on my XTX vs the Pulse, I just wanted to not have to mess with it down the road as I keep my GPUs for a bit.

According to Sapphire's product pages, the 9070 PULSE cards now both come with Honeywell.

 

[itsmydamnation]

While that is generally true, we shouldn't wait for drivers to mature. AMD is a multi billion Dollar company, if drivers are garbage again than that is close to unacceptable. What you wrote basically sounds like we are at a point where it is acceptable for AMDs drivers to be bad at the beginning and it's great when they mature sometime in the future. However that is not what we should expect from a company like that.

Also looking at RDNA2: 6900XT had 11% more CUs than 6800XT without any other changes: Same clocks, same TDP. Yet it delivered like 8-9% perf at 4K over 6800XT at that same TDP. So RDNA4 scaling seems to be quite meh for now, we will see how N44 turns out later this year.

For now it could be a first indication why other RDNA4 was cancelled, at least seems more likely than just "GB202 too strong, we give up".

do you people even look at specs before going batty?

they have the exact same memory speed and LLC size, we already know this is going to be a bandwidth staved card is has ~ 7900XTX performance with 2/3rd the LLC and 2/3 the memory bandwidth.

 

[adroc_thurston]

6900XT had 11% more CUs than 6800XT without any other changes: Same clocks, same TDP. Yet it delivered like 8-9% perf at 4K over 6800XT at that same TDP.

Not really.
Remember the quirk: for quite a lot of cases (particularly load-balanced PS waves) 6800XT behaves like a 64CU GPU.

 

[Mahboi]

From reddit and techpowerup:

The new Radiance 2 Display Engine comes with major hardware updates that reduce GPU idle power draw in multi-monitor setups. The engine also comes with hardware flip-metering support (something NVIDIA also introduced with Blackwell, and which enables Multi-Frame Gen on the RTX 50-series). Flip-metering improves video frame pacing to the GPU and reduces CPU overhead for video playback. There is also a display engine level hardware image sharpening component that drives Radeon Image Sharpening. As for I/O, you get contemporary DisplayPort 2.1a and HDMI 2.1b, we are waiting to confirm the UHBR bit-rates.

Navi 48 comes a dual media engine, each of the two can perform concurrent encoding and decoding, and so the implementation resembles an NVIDIA GPU having two each of NVENC and NVDEC units. The new generation media engine offers a 25% increase in H.264 low-latency encode quality, and an 11% improvement in HEVC encode quality. AV1 encode and decode gets B-frames support, vastly improving bitrates. A B-frame is a frame that lacks image information, but instead motion-vector values, that lets the decoder reconstruct the image component using past and future I-Frames (frames with image data). The media engine posts a 50% generational performance uplift (measured in encoder/decoder frame rates), with reductions in memory overhead.

I WAS trying to not spend money on a new GPU when RDNA 3 is enough except for the idle and media consumption power draw.
STOP MAKING ME WANT TO BUY AYYYMD

 

[APU_Fusion]

I think AMD should have set the price at 49.99 and 99.99 myself /s

 

[gdansk]

I think AMD should have set the price at 49.99 and 99.99 myself /s

Okay at those prices no one could complain.

But reading the TPU overview they mentioned something about $350 which I consider about as delusional unless they were talking about a discounted 3/4th chop.

Should there be a price-war with the RTX 5070 series, we can expect AMD to drag prices all the way down to below $350.

 

[adroc_thurston]

unless they were talking about a discounted 3/4th chop.

the problem with 3/4th chops is TSM being the king of yield.

 

[SolidQ]

Youtube! Leaked, not licked!

 

[gaav87]

Chinese can already pre order 9070xt on JD for past 8h
Most models: pulse, steel legend, prime - msrp
sapphire pure +50$
Red devil +150$
xfx mercury +75$ magnetic air +100$ oc version +120$

I can go to sleep now and check in 1 week thank god.

 

[Mahboi]

I WAS trying to not spend money on a new GPU when RDNA 3 is enough except for the idle and media consumption power draw.
STOP MAKING ME WANT TO BUY AYYYMD

Oh wow I'M SAVED

Spoiler

Not a single model of the new GPUs will fit in my tiny 315mm length mATX case
yayyyy.............................................................................

 

[Mahboi]

Past my silly posts, I do like the huge effort where AMD was really lacking: media consumption, RT, better idle power draw...there's really a lot to like, all the little things that they were behind have been shored up seemingly.
My only 3 metrics left are:
- VR support speed (took SEVEN MONTHS for RDNA 3 to function like RDNA 2 did)
- ROCm support speed (Phoronix is already lamenting that months is too much, and it will be months)
- Broader RDNA 4 features support in 3rd party elements

For the latter, I'm really hopeful that some of the most well-known weaknesses in say Blender actually get pounded away, and that several of the other titles get support before 3 months have to pass. Premiere Pro, Photoshop, etc...