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

 

[adroc_thurston]

Then why is it arch, 32 previous name navi ? and arch, 64 is MI ?

Baseline wave size.

 

[gaav87]

Baseline wave size.

So next MI is simd32 with 2 simd per CU and 2 CU per WGP just like rdna ?

OR 1250 and 1251 are BOTH MI and CONSUMER !

 

[adroc_thurston]

So next MI is simd32 with 2 simd per CU and 2 CU per WGP just like rdna ?

Yeah.

OR 1250 and 1251 are BOTH MI and CONSUMER

The next client gfx is 13.

 

[gaav87]

Yeah.

The next client gfx is 13.

Well i think you wrong as MI300 did not use WGP but XCD and wgp was never used in documentation.
I just wanted to know if you knew xD

 

[soresu]

It's MI450X.

Are they skipping MI400 naming to MI450?

 

[adroc_thurston]

Are they skipping MI400 naming to MI450?

They name it whatever.

 

[marees]

Are they skipping MI400 naming to MI450?

MI400 next year
MI500 the year after

Also I think they could do an AI card on RDNA 4 (& GDDR 7) by end of this year or early next year as speculated (wished for) by MLID

 

[soresu]

Phoronix was more thorough than I documenting it...

AMD Previews Instinct MI400 Series & Helios AI Rack - Phoronix

www.phoronix.com

 

[soresu]

Phoronix's ROCm 7 post:

ROCm 7.0 Goes Into Preview With MI350X/MI355X Support, Big Performance Improvements - Phoronix

www.phoronix.com

 

[soresu]

Also I think they could do an AI card on RDNA 4 (& GDDR 7) by end of this year or early next year as speculated (wished for) by MLID

Not gonna happen, there's no call for an AI SKU for RDNA4 when MIxxx exists.

Larger and faster VRAM can benefit a host of professional gfx (DCC) use cases though.

 

[eek2121]

In theory Radeon is improving. RDNA is a good architecture, however is AMD really closing in the gap?

Once again there's discussion and angry comments that Nvidia is selling less for a higher price:

View attachment 125211

But seeing these number I realise, each generation renamers a lower tier die as a higher tier card and yet it's enough to compete with AMD.
If Jensen wanted he could really have "killed" Radeon, but the way it is AMD's GPUs are no more than another tool Jensen uses to boost the value of his products and brand.


It's depressing.

I somehow missed this and I had to comment. NVIDIA is doing this because their costs have shot up drastically the past few years. Moving from Samsung to TSMC means they have to pay quite a bit more per wafer. They had some pretty sweet discounts at Samsung. No they could not 'kill' Radeon. They would have to sacrifice margin, and they already think they need a higher margin than they have. If NVIDIA pushed out a bigger part, AMD has cards they can play. They do have higher CU designs in house that they can develop out (the "canceled" NAVI4c/etc). However, AMD wants to pursue clocks instead. Higher clocks on smaller dies, which I think is the right approach. They are getting there.

Both companies are also actively pursuing new upscaling initiatives. Both companies hope to make upscaling 'as good as' the real thing so they can get away with smaller dies. I suspect NVIDIA may flop hard on this and try to do it too soon. Hopefully AMD is smart and waits until the tech is perfect.

Newer processes also have costs increasing more than the die is shrinking, so the cost per die is going up as newer and newer nodes are used.

 

[reb0rn]

We see with clock fight, Nvidia today is way more efficient if you take out cut down 9070, if they pump clock more they efficiency vs NV will drop even more down and then ppl will start to see it more and more same as all AMD GPU use more TDP then declared compared to NV and that is fact for last 10y, so good luck gimping NV with clock

 

[gdansk]

"If you remove their highest performance per watt product from this generation then they are behind"

Absolutely brilliant analysis, thank you Einst3in

 

[Thunder 57]

However, AMD wants to pursue clocks instead. Higher clocks on smaller dies, which I think is the right approach. They are getting there.

...

Newer processes also have costs increasing more than the die is shrinking, so the cost per die is going up as newer and newer nodes are used.

Wider designs with lower clocks have proven to be more efficient time and time again. The 9000 series loses in efficiency because they relied to heavily on clocks. You may be right that is the way to go in the future with larger dies costing more and more.

 

[soresu]

We see with clock fight, Nvidia today is way more efficient if you take out cut down 9070, if they pump clock more they efficiency vs NV will drop even more down and then ppl will start to see it more and more same as all AMD GPU use more TDP then declared compared to NV and that is fact for last 10y, so good luck gimping NV with clock

AMD's high end SKUs always chase the less efficient end of the bell curve to squeeze the last bit of performance from smaller dies.

Saying "exclude the SKU that doesn't sacrifice efficiency for max perf" when talking about efficiency is a fundamentally flawed take.

 

[soresu]

However, AMD wants to pursue clocks instead. Higher clocks on smaller dies, which I think is the right approach. They are getting there.

Hypothetically this is the perfect strategy if you make the assumption that gfx chiplets was always their endgame.

Once they have chiplets perfected they can spam them at lower clocks a la Epyc and win the perf/watt competition with ease while still remaining competitive in absolute perf.

 

[eek2121]

Wider designs with lower clocks have proven to be more efficient time and time again. The 9000 series loses in efficiency because they relied to heavily on clocks. You may be right that is the way to go in the future with larger dies costing more and more.

Wider designs are becoming unaffordable, unfortunately. For GPUs at least…

Hypothetically this is the perfect strategy if you make the assumption that gfx chiplets was always their endgame.

Once they have chiplets perfected they can spam them at lower clocks a la Epyc and win the perf/watt competition with ease while still remaining competitive in absolute perf.

Chiplet plans are shelved for now. I am sure somebody at AMD is still trying to get it working, however, as I understand it, no gaming products planned. That could always change, of course.

AMD is actually very good at keeping secrets.

 

[Thunder 57]

Hypothetically this is the perfect strategy if you make the assumption that gfx chiplets was always their endgame.

Once they have chiplets perfected they can spam them at lower clocks a la Epyc and win the perf/watt competition with ease while still remaining competitive in absolute perf.

So instead of "E core spam" we will have "R die spam"?

 

[soresu]

So instead of "E core spam" we will have "R die spam"?

"SED spam" technically going by terminology that was floating around some time back, might have been from AMD patents on GPU chiplets.

 

[soresu]

AMD is actually very good at keeping secrets.

In the worlds of random GoT person:

 

[GaiaHunter]

Overclockers.co.uk has the Powercolor 9070xt hellhound for £629, https://www.overclockers.co.uk/powe...-pci-express-graphics-card-gra-pwc-04016.html

That is the lowest I've seen the 9070xt since launch. They also have the asrock steel legend for £650. That is what I paid for mine at launch.

The 9070 Reaper is going for £560 and the 9070 Pulse for £570.

 

[soresu]

I can only dream of the day when AMD offer a better performing card for the same price as Nvidia.

Either they will offer an equal perf card at a lower price to gain market share, or they will offer a superior card that costs more because they are the front runner.

Once you have a superior product you can demand top dollar.

That's just the way industry works.