Discussion RDNA4 + CDNA3 Architectures Thread

DisEnchantment · Mar 23, 2022

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 · Monday at 12:07 PM

marees said:
But the steam deck (this now sells half as much as xbox) could go nvidia+arm

Doing double emulation (DirectX to Vulkan + x86 to ARM) is just asking for trouble.

Ajay · Monday at 12:11 PM

leoneazzurro said:
Not saying that is not important there, but IIRC you need even more in a gaming GPU due to the presence of other requirements (i.e texturing) and anyway a MI300X-like packaging should still cost a lot for consumer graphics.

That sort of packaging is far too expensive for consumer graphics - and the advanced packaging facilities are already completely booked out.

blckgrffn · Monday at 12:31 PM

Ajay said:
That sort of packaging is far too expensive for consumer graphics - and the advanced packaging facilities are already completely booked out.

The real story. AI sucking all the oxygen out of the room.

inquiss · Monday at 12:51 PM

jpiniero said:
Good catch. So def fake.

Or there is no halo anymore. Ramping AI has taken some pain?

jpiniero · Monday at 12:59 PM

inquiss said:
Or there is no halo anymore. Ramping AI has taken some pain?

I don't expect AMD to challenge in the high end again until they get chiplets working. I suppose it's possible they still are having problems.

inquiss · Monday at 1:54 PM

jpiniero said:
I don't expect AMD to challenge in the high end again until they get chiplets working. I suppose it's possible they still are having problem

Even then, there might be better ways to get return on that investment than a high end graphics card no one will buy. While AI is printing this much money, why spend the moneys en time on the halo when you can get AI-bux for that capacity and engineering talent.

blckgrffn · Monday at 2:16 PM

inquiss said:
Even then, there might be better ways to get return on that investment than a high end graphics card no one will buy. While AI is printing this much money, why spend the moneys en time on the halo when you can get AI-bux for that capacity and engineering talent.

When you can sell every mega buck AI accelerator you can produce the entire consumer GPU segment is something you put on maintenance support. It's your duty to your shareholders to put your best talent on that money printing and extremely competitive and time sensitive front and push the pedal to medal to ride the tiger for as long as possible.

It's not on topic for this thread but I fully expect nvidia to continue to embrace this because I think they are already pivoted this way.

As you stated, there are many shades of gray between things like full support and life support and we may see efforts wax and wane as time goes on but I really expect this to be the case.

Ajay · Monday at 2:32 PM

blckgrffn said:
The real story. AI sucking all the oxygen out of the room.

And at AMD it literally sucked engineers out of the consumer GPU development team.

blckgrffn · Monday at 2:47 PM

Ajay said:
And at AMD it literally sucked engineers out of the consumer GPU development team.

I mean, they gotta do it. It doesn't make me happy but I get it.

beginner99 · Tuesday at 1:23 AM

blckgrffn said:
It's not on topic for this thread but I fully expect nvidia to continue to embrace this because I think they are already pivoted this way.

sadly I agree and that is why nothing will change for next generation. a very minor performance/$ increase, barley more vram, in essence not much at all. There is zero, zero reason for either company to price their GPUs cheaply. the factories are full pumping out AI cards, no need to enter a price war to fill the factories so better to increase margin.
Intel was the only hope but since battlemage seems to come from TSMC as well, there is the same issue, no need to fill their own factories and enter a price war.

Saylick · Tuesday at 1:32 PM

Chips & Cheese tested MI300X. Really interesting plots. Looks like they tested cache latency, cache bandwidth, compute throughput, link bandwidth between dies, and some benchmarking. Enjoy!

Testing AMD’s Giant MI300X

AMD’s Radeon Instinct MI300X is the latest in the company’s compute focused CDNA line. NVIDIA has dominated the GPU compute market for time immemorial, thanks to a combination of strong…

chipsandcheese.com

adroc_thurston · Tuesday at 1:38 PM

marees said:
But the steam deck (this now sells half as much as xbox) could go nvidia+arm

Because all the previous Tegra handhelds went great right?

eek2121 · Tuesday at 4:16 PM

marees said:
At this point I am just hoping that AMD doesn't lose the surface handheld to Intel

The Xbox next seems locked to RDNA 5
& PlayStation seems locked to Radeon for a long time
But the steam deck (this now sells half as much as xbox) could go nvidia+arm

The Steam Deck is modeled after hardware found in the consoles and PC (well they try), so it won’t ever see ARM unless the consoles do. They will likely stick with an AMD GPU for similar reasons.

JustViewing · Wednesday at 1:46 AM

MI300X is a beast . Maybe in 4/5 years we can have similar power in desktop GPUs

igor_kavinski · Wednesday at 5:40 AM

JustViewing said:
MI300X is a beast . Maybe in 4/5 years we can have similar power in desktop GPUs

Maybe we can get them cheap on ebay by that time

darkswordsman17 · Wednesday at 7:43 AM

adroc_thurston said:
Because all the previous Tegra handhelds went great right?

Its done well for Nintendo, and developers seem to like it (which means there's already a lot of software resources ready). The Nvidia GPU drivers alone would likely fix most of the issues that Qualcomm's chips on Windows are having. They'd likely be able to ask as much as Qualcomm if not more so their greed won't even be a factor since it'll let OEMs shave a bunch of other costs.

eek2121 said:
The Steam Deck is modeled after hardware found in the consoles and PC (well they try), so it won’t ever see ARM unless the consoles do. They will likely stick with an AMD GPU for similar reasons.

Did you forget the Switch exists?

adroc_thurston · Wednesday at 11:22 AM

darkswordsman17 said:
Its done well for Nintendo,

But so did Wii.

igor_kavinski · Wednesday at 12:12 PM

adroc_thurston said:
But so did Wii.

I do want to know what soured Nintendo away from AMD. Lisa is usually very accommodating to partnerships and existing partners.

Hitman928 · Wednesday at 12:30 PM

igor_kavinski said:
I do want to know what soured Nintendo away from AMD. Lisa is usually very accommodating to partnerships and existing partners.

NV had a cheap, mobile oriented SoC ready to go. AMD did not and was not in a place to expand their offerings when the Switch was in development without major up front costs being handled by Nintendo (which they would never do). So, I think the choice was pretty simple for them.

GodisanAtheist · Wednesday at 2:43 PM

Hitman928 said:
NV had a cheap, mobile oriented SoC ready to go. AMD did not and was not in a place to expand their offerings when the Switch was in development without major up front costs being handled by Nintendo (which they would never do). So, I think the choice was pretty simple for them.

- Heard it here first folks, Nvidia is actually responsible for the Steam Deck and AMD's handheld SOC success.

All roads lead back to Nvidia

soresu · Wednesday at 3:14 PM

Hitman928 said:
NV had a cheap, mobile oriented SoC ready to go. AMD did not and was not in a place to expand their offerings when the Switch was in development without major up front costs being handled by Nintendo (which they would never do). So, I think the choice was pretty simple for them.

I wanna say Kabini?

Not ideal power consumption wise for sure, but then neither is TX1 at full power.

Plus Kabini would have allowed them a significant level of code compatibility with PS4 and XB1.

IMHO Ninty simply went with TX1 because nVidia offered them a dirt cheap price

igor_kavinski · Wednesday at 3:25 PM

soresu said:
IMHO Ninty simply went with TX1 because nVidia offered them a dirt cheap price

Or an employee had a Tegra device which they rooted and somehow ported Mario over to it and then showed it to their bosses and they went "Mamma Mia!".

Hitman928 · Wednesday at 3:31 PM

soresu said:
I wanna say Kabini?

Not ideal power consumption wise for sure, but then neither is TX1 at full power.

Plus Kabini would have allowed them a significant level of code compatibility with PS4 and XB1.

IMHO Ninty simply went with TX1 because nVidia offered them a dirt cheap price

Kabini was too old and would have had terrible performance in the power envelope they needed.

Glo. · Wednesday at 3:41 PM

Saylick said:
Chips & Cheese tested MI300X. Really interesting plots. Looks like they tested cache latency, cache bandwidth, compute throughput, link bandwidth between dies, and some benchmarking. Enjoy!

Testing AMD’s Giant MI300X

AMD’s Radeon Instinct MI300X is the latest in the company’s compute focused CDNA line. NVIDIA has dominated the GPU compute market for time immemorial, thanks to a combination of strong…

chipsandcheese.com

This is a testament for unified memory architectures.

Finally people can see the light .

soresu · Wednesday at 3:44 PM

Hitman928 said:
Kabini was too old and would have had terrible performance in the power envelope they needed.

How quickly people forget that A15/57 were power hog µArchs.

IMHO Maxwell and GCN2 were also fairly comparable in feature set sans the double rate FP16 that the TX1's GPU had.

TX1 was fabbed on TSMC 20nm vs Kabini's 28nm, but that was felt to be a poor increment over 28nm from what I remember at the time - such that neither AMD nor nVidia actually made any discrete GPUs on the node.

Edit: Seems like Kabini/Mullins GPU was only 2 CU, I thought it was twice that.

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Senior member

Lifer

Diamond Member

Member

Lifer

Member

Diamond Member

Lifer

Diamond Member

Diamond Member

Diamond Member

Diamond Member

Diamond Member

Member

Lifer

Lifer

Diamond Member

Lifer

Diamond Member

Diamond Member

Platinum Member

Lifer

Diamond Member

Diamond Member

Platinum Member