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

 

[Kepler_L2]

The mofos at wtftech are quick… JFC:
View attachment 129604

Also, customary “Hassan, go f yourself”

I'll have to start posting absolute nonsense from time to time just to keep the slop farms on their toes

 

[Saylick]

I'll have to start posting absolute nonsense from time to time just to keep the slop farms on their toes

You just gotta embed some insults to them as part of your post, so that if they screenshot your post, they’d only be broadcasting it out.

 

[adroc_thurston]

The giant bump in L2 offsets the area loss of MALL pretty heavily, so you must be talking about the shader core.

They made L2 denser doe.

Still better than now, or worse, it depends.
If 8SE/96WGP equivalent is that far off the top then the rest of the stack is doomed no matter what.
Like you really expecting Rubin to be another Ampere>Ada leap? With a much less significant node shrink?

Oh noes, just on top.

Giving up" would have meant taking away all the ROP blocks out of AT0, making it unusable for gaming

It's for cloud gaming.
You can't chop off ROPs there.

They did not and AT0 is coming to the DIY market, so they're not giving up.

A heavily castrated chop that won't be competitive.

 

[Magras00]

gfx13 codenames are from Transformers

Alpha Trion, (Ultra) Magnus, Orion (Pax)

Is PS6 Canis based on Canis Minor-8?

Fandom page says beast mode is a puppy.

 

[Thunder 57]

You can't market a product that doesn't win.

Either they ship 320CUs@800W and win, or they lose.
Pretty dang simple!

800W would be stupid. Nobody wants that.

 

[adroc_thurston]

800W would be stupid. Nobody wants that.

Stupid? Yes.
But everybody wants a winner.

 

[marees]

Is PS6 Canis based on Canis Minor-8?

Fandom page says beast mode is a puppy.

Both Orion & Canis are related constellations

Orion, The Celestial Centurion and His Faithful Companions » Astronomy For Change

astronomyforchange.org

 

[Win2012R2]

800W would be stupid. Nobody wants that.

Will sell like very hot cakes if it allows to reach 100 fps for UE5 games at 4k.

 

[ToTTenTranz]

Will sell like very hot cakes if it allows to reach 100 fps for UE5 games at 4k.

Now now, let's not be too hasty here.
75FPS 4K on UE5 games. With upscaling.

 

[branch_suggestion]

They made L2 denser doe.

That was a given, just being N3P helps densify the macro but obv needed to be closer to MALL density for PPA.
Oh and the area calc is based on AT2 being supposedly 264mm^2, most areas would be 2.4x, others <2x-2.66x.
So roughly 600mm^2.

Oh noes, just on top.

Then your perf projection for AT0 is lower than mine.

 

[adroc_thurston]

So roughly 600mm^2.

Pretty sure less.

Then your perf projection for AT0 is lower than mine.

It's less about perf and more about comp projections.
AMD doesn't have a single sliver of chance on top without resorting to SoIC. Which they won't.

 

[Win2012R2]

Now now, let's not be too hasty here.
75FPS 4K on UE5 games. With upscaling.

SOLD!

 

[branch_suggestion]

Pretty sure less.

Will probably end up at 550ish through torturing RTL engineers.

It's less about perf and more about comp projections.
AMD doesn't have a single sliver of chance on top without resorting to SoIC. Which they won't.

I do agree, NV can beat AT0 with their known capabilities, just that they might need to get closer to the ret limit then they would like.
A 2x ret tiled thing on the other hand, no chance.
But better to wait for a special node to try that on, like A16.

 

[adroc_thurston]

Will probably end up at 550ish through torturing RTL engineers.

We love that don't we folks.

NV can beat AT0 with their known capabilities, just that they might need to get closer to the ret limit then they would like.

that's the point, AMD's not gonna win.

But better to wait for a special node to try that on, like A16.

A16 is a dead node for 0 (zero) customers.
Not like it matters anyway. AMD's not competing with Nvidia in discrete graphics anymore.

 

[Win2012R2]

New stuff better be reticle sized limit because Nvidia's will be.

 

[Thunder 57]

Will sell like very hot cakes if it allows to reach 100 fps for UE5 games at 4k.

Oh they'll be hot alright. Space heater hot.

 

[adroc_thurston]

Oh they'll be hot alright. Space heater hot.

Doesn't matter as long as you win.

 

[branch_suggestion]

Not like it matters anyway. AMD's not competing with Nvidia in discrete graphics anymore.

You could say that ever since N36 got shelved.
RDNA3 really needed to not be made of suck.
If there is a question to be asked, is it inherently bad that they are pivoting away from discrete graphics comp to GPGPU and APUs?
It's not like they are taking resources away from graphics, the IP is progressing nicely and they are getting the bulk of the coolest parts over the CPU division.

 

[adroc_thurston]

You could say that ever since N36 got shelved.

no, when N4c got axed.

it is inherently bad that they are pivoting away from discrete graphics comp to GPGPU and APUs?

It's just boring as hell.

It's not like they are taking resources away from graphics, the IP is progressing nicely and they are getting the bulk of the coolest parts over the CPU division.

It's good for what I actually shop for (laptops) and bad for the fun factor since you have good IP wasted on configs that don't take any heads.
Like in GPGPU they're on a jihad to make the meanest, baddest GPU socket, rack and interconnect? Client dGFX? you get scraps off someone's table.

 

[branch_suggestion]

It's just boring as hell.

The market is boring as hell, NV really launched a very meh generation.
They get rewarded with their best client q ever. Market has been Intel-fied, let it rot.

 

[adroc_thurston]

The market is boring as hell, NV really launched a very meh generation.

They still won and that is all that matters.

They get rewarded with their best client q ever. Market has been Intel-fied, let it rot.

Yeah that's called winning.
Intel was also winning.

 

[Darkmont]

Dunno how the market could ever escape the Jensen cult outside of at least 3-4 sustained ass beatings from some Radeon chiplet battle axe thing

 

[adroc_thurston]

Dunno how the market could ever escape the Jensen cult outside of at least 3-4 sustained ass bearings but some chiplet battle axe thing

well, duh.
either AMD embraces the Ork warboss lifestyle and builds the biggest choppa' again and again, or it is stuck forever.

worked every bit as well as one could imagine in DYI DT CPUs so they just have to commit.

 

[branch_suggestion]

They still won and that is all that matters.

Yeah that's called winning.
Intel was also winning.

N4C could've won...
Or it could've been a catastrophe, we shall never know in this timeline.
All we do know is N48 is actually alright, all it took was a good enough upscaler.

 

[adroc_thurston]

All we do know is N48 is actually alright, all it took was a good enough upscaler.

Yeah, decent part sold for healthy ASPs, but it just gives AMD zero notoriety.