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

 

[Win2012R2]

I'm still waiting for disappointment™.

Availability ...

 

[adroc_thurston]

RDNA is cool again, they'd be very stupid to change away from RDNA5

Change is inevitable and so is gfx13

 

[poke01]

I'm still waiting for disappointment™.

But perhaps for the first time in a long time I might be disappointed to not be disappointed?

Well apparently, in Australia the distributions started pre-scalping so retailers have to start range from at $1200 to $1500AUD.

But the stock levels are good much better than NVs

 

[adroc_thurston]

Well apparently, in Australia the distributions started pre-scalping so retailers have to start range from at $1200 to $1500AUD.

But the stock levels are good much better than NVs

Is the demand level really that insane

 

[jpiniero]

Is the demand level really that insane

Given that there's like no 50 Series supply, it's kind of the only game in town.

 

[RnR_au]

Is the demand level really that insane

I think there is a bit of FOMO going on. One of the main retailers here in AU posted that from a stock point of view, this is one of the best gpu releases ever. But they also said that stock might not last one day and they had buffed their servers.

 

[H T C]

What are the odds AMD releases "new drivers" a bit after launch that boost the 9000 series by "a significant %", in order to catch nVidia "with their pants down"?

IMO, the odds are EXTREMELY LOW ... but i wouldn't mind "being disappointed" if they weren't ...

 

[Tuna-Fish]

What are the odds AMD releases "new drivers" a bit after launch that boost the 9000 series by "a significant %", in order to catch nVidia "with their pants down"?

IMO, the odds are EXTREMELY LOW ... but i wouldn't mind "being disappointed" if they weren't ...

Intentionally? Zero. That would be the dumbest move ever. The launch reviews matter for the entire product life cycle. When someone considers what GPU to get 6 months from now, they are much more likely to see a comparison article or video that got a lot of links and clicks at launch time (and so is still high up in search engine results), than they are to see a "how does it perform today" article. We saw how destructive this was with the AMD "finewine" drivers (read: they didn't get drivers properly done until way after launch), they were toxic to mindshare and perception.

Still can happen if they fix issues and find bugs after launch.

 

[RnR_au]

From an AI perspective...



Source

 

[PJVol]

From an AI perspective...

Can AI "mine" $$ ?

 

[ToTTenTranz]

From an AI perspective...

View attachment 118546

Source

This also puts RDNA4's per-CU INT8 throughput ahead of the PS5 Pro's GPU by a large margin.
If the console does ~300 TOPS INT8 with a 2.35GHz GPU then it's kind of a mid way between RDNA3 and 4.

 

[branch_suggestion]

What are the odds AMD releases "new drivers" a bit after launch that boost the 9000 series by "a significant %", in order to catch nVidia "with their pants down"?

IMO, the odds are EXTREMELY LOW ... but i wouldn't mind "being disappointed" if they weren't ...

Other than an enforced memoc on the XT, not much can be done.
Vanilla 9070 I could certainly see a driver update to a 260W TBP to counter a 5070 Super or whatever in tandem with a price cut.
The current config is built to win the efficiency graphs, nothing more.

 

[H T C]

Vanilla 9070 I could certainly see a driver update to a 260W TBP to counter a 5070 Super or whatever in tandem with a price cut.

When i said "a significant %", i was thinking along the lines of 15%+ overall: enough to overtake whatever card's performance nVidia will have closest to it @ launch, but WITHOUT resorting to changing the wattage to do it.

 

[RnR_au]

Can AI "mine" $$ ?

I don't think it has the needed memory bw. Crypto miners are offloading the Instinct 50's and 60's and they have 1TB/s membw. But I have never been in the crypto space... so kinda clueless.

 

[Abwx]

Other than an enforced memoc on the XT, not much can be done.
Vanilla 9070 I could certainly see a driver update to a 260W TBP to counter a 5070 Super or whatever in tandem with a price cut.
The current config is built to win the efficiency graphs, nothing more.

No need to release any driver, that s already done by some brands, for both SKUs btw.

Nitro+, Pure und Pulse: Sapphires Radeon RX 9070 (XT) im Überblick

Mit sechs Modellen startet Sapphire in die neue Radeon-Generation. Jeweils gibt es die RX 9070 XT und RX 9070 als Nitro+, Pure und Pulse.

www.computerbase.de

 

[itsmydamnation]

man...... if only AMD had a 96CU // 384bit part ... my 7900XTX is 6 months olds , i'll still probably buy a 9090XTX. But right now its like RDNA 4 can do lower precision better , but has less memory (24gb vs 16gb matters for local LLM's ) and doesn't perform any better until FSR4 makes XESS on RDNA3 look like crap.

the distilled DS models would love RDNA4 , RDNA3 // 7900XTX was already at 4090 levels of performance in it.

 

[PJVol]

No need to release any driver, that s already done by some brands, for both SKUs btw.

Looks like both gpus hosted the same pcb, otherwise how to explain the need in "12V-6+2-4*8 " (or whatever) power connector for the 250W gpu?

 

[Abwx]

There s no pic of the 245W 9070 mounting block, what we can see is that the Nitro+ has 6 pipes cooler while the 304/317W 9070XTs have the same 5 pipe cooler, the 220W 9070 has 4 pipes one, methink that the 245W version has te same as the standard 9070XT.

Beside the power connector is placed differently in the Nitro+ and the Pure, somewhat weirdly for the former, the 9070s have the power connector at the usual place, so there s at least two different PCBs.

 

[poke01]

AMD RDNA4 Architecture Speculation

According to a former Apple graphics enginer, a closer description to their "dynamic caching" technology would be dynamic 'deallocation' which allows them to release unified/flexible on-chip memory during runtime depending on whichever path of execution or branch is taking place within a shader...

forum.beyond3d.com

This page goes in depth regarding RDNA4 RT.

 

[adroc_thurston]

This page goes in depth regarding RDNA4 RT.

It's a very funny implementation since it does the exact opposite of what everyone else is doing.

 

[Win2012R2]

It's a very funny implementation since it does the exact opposite of what everyone else is doing.

Is it better?

Can they at least double RT perf on RDNA5?

 

[adroc_thurston]

Is it better?

different.
Raytracing as a workload is a bespoke kind of horror for GPUs: it's very divergent and very very latency-sensitive, and GPUs naturally suck at that.
Which is why everyone else bolted a purpose-build coprocessor that offloads the tricky bits of the RT pipeline away from the shader cores (https://www.intel.com/content/www/us/en/developer/articles/guide/real-time-ray-tracing-in-games.html as an example. Intel does implicit h/w thread sorting, even).

AMD went the opposite, pure MONKE GRUG SMASH way and addressed the problem head on: their new shader core handles tricky waves flexibly and hides latency a lot lot better since the ld/st unit has (however limited) reordering capabilies.
You don't need a MIMD walker when you can just hide the latency away and who needs coherence sorting anyway just spill the regs whatever.

Can they at least double RT perf on RDNA5?

Maybe?
If they're really building an OoO shader core, pretty much anything is possible.
A P6 moment for GPUs would be insane.

 

[Win2012R2]

their new shader core handles tricky waves flexibly

Yeah I like that very much - it makes them more useful for other possible workloads that would otherwise suck, but the question is - in order to scale perf do they also have to scale number of CUs or can they double smaller relevant bits inside CU to double specifically RT perf?

 

[adroc_thurston]

in order to scale perf do they also have to scale number of CUs or can they double smaller relevant bits inside CU to double specifically RT perf?

if gfx13 really goes OoO, I don't even know.
A completely different paradigm, nothing else would be even remotely similar.

 

[Win2012R2]

Have they improved dual issue to be at least on par with Nvidia's?