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

beginner99 · May 21, 2024

adroc_thurston said:
you dropped this banger the day MS actually refreshed it with relevant h/w for once.

Not being x86 makes my argument to buy a mac even more relevant? If I wanted max efficiency at good performance I would buy a mac. No debate they are the best in that regard. Being on Windows and not x86, I don't know. Would only work for me as a secondary device which I'm not paying $2000 for reasonable hardware in terms of RAM and SSD.

adroc_thurston · May 21, 2024

beginner99 said:
If I wanted max efficiency at good performance I would buy a mac.

Hamoa laptops do that just fine.

ToTTenTranz · May 21, 2024

adroc_thurston said:
Hamoa laptops do that just fine.

How high is the compatibility with x86 apps though?

adroc_thurston · May 21, 2024

ToTTenTranz said:
How high is the compatibility with x86 apps though?

Everything non AVX will be aight.

ToTTenTranz · May 21, 2024

adroc_thurston said:
Everything non AVX will be aight.

This leaves out a bunch of stuff, though. Video editing and transcoders, for example.

adroc_thurston · May 21, 2024

ToTTenTranz said:
Video editing

Offloaded to a FF block.

ToTTenTranz said:
transcoders

Ditto.

ToTTenTranz · May 21, 2024

adroc_thurston said:
Offloaded to a FF block.

Ditto.

But that requires support in the software, probably a new license and all.

I've been doing basic video editing for powerpoint presentations with the same Vegas 14 license I bought from humble bundle like 10 years ago. I'd guess many people will be in a similar circumstance.

Mopetar · May 21, 2024

adroc_thurston said:
it is what it is, 5 years and RTRT hasn't left the gimmick stage.

People will stop caring about it when the next gimmick is revealed. I'm guessing that the next console generation will talk about 8K gaming, and suddenly that will be the new thing everyone is talking about. This generation it was RT, last generation was VR, before that was motion controls. Before that it was HD gaming.

I think the next generation could be 8K gaming. By then the consoles will be close to good enough to do native 4K and they can bake in some upsampling hardware for the people who want to show off their shiny new 8K displays.

gdansk · May 21, 2024

Mopetar said:
People will stop caring about it when the next gimmick is revealed

Doubt it. They'll care about it until AMD/Intel are just as good. The next gimmick will only become a deal breaker when old gimmicks are not.

Aapje · May 21, 2024

gdansk said:
They'll care about it until AMD/Intel are just as good.

This makes no sense. Most people have Nvidia anyway.

The issue is that Nvidia for the most part is also not good enough with RT.

gdansk · May 21, 2024

Aapje said:
This makes no sense. Most people have Nvidia anyway.

The issue is that Nvidia for the most part is also not good enough with RT.

Not the point. The mere reveal doesn't do much. There's almost always a generation or two before a new gimmick becomes unused (3D displays), is matched (tessellation) or it becomes a dealbreaker (DLSS).

And in the first two cases people stop caring quickly. But people overlook Radeon now because they haven't caught up in RT or upscaling.

inquiss · May 21, 2024

gdansk said:
Not the point. The mere reveal doesn't do much. There's almost always a generation or two before a new gimmick becomes unused (3D displays), is matched (tessellation) or it becomes a dealbreaker (DLSS).

And in the first two cases people stop caring quickly. But people overlook Radeon now because they haven't caught up in RT or upscaling.

Nvidia is always better because of "The Latest Thing". Doesn't matter what it is or if it's usable or any good. It's "The Latest Thing", and AMD doesn't have that even if it doesn't matter. Buy Nvidia.

Shmee · May 21, 2024

The latest things these days are often dumb and for dumb kids these days. And yes I am grumpy.

Kepler_L2 · May 21, 2024

inquiss said:
Nvidia is always better because of "The Latest Thing". Doesn't matter what it is or if it's usable or any good. It's "The Latest Thing", and AMD doesn't have that even if it doesn't matter. Buy Nvidia.

Unless "The Latest Thing" is superior DX12/Vulkan performance, FP16 or real Async Compute hardware. Then it doesn't matter

soresu · May 21, 2024

ToTTenTranz said:
Current solutions consist of brute-forcing too much which then brings $2500 GPUs to their knees.

But for Nvidia that's okay because that $2500 GPU is brought to its knees a bit less than the competition and Nvidia's previous $2500 GPU

IMHO for nVidia it's more than OK, because the entire point in pushing RTRT was raster graphics weren't enough to strain GPUs enough anymore at 4K in the high end.

Resolution scaling solutions to this problem like 8K were a non starter for people that don't have a sizable room to play with, the alternative being VR which was still extremely niche at the time they made the decisions, even now it's just not high enough volume to save raster gfx.

Effectively nVidia used RTRT to move the performance target goal posts - without which they were likely worried that consumer GPU sales might have cratered by now, if not in the very near future - certainly without putting time into RTRT code development we can expect that an extra 6 years of raster gfx research would have yielded even more features and perf improvements.

Getting to one up AMD who weren't as focused on hardware RT was just a bonus considering nVidia were already the market leader when Turing dropped.

That being said, in order to continue pushing RTRT with a straight face they have had to research their backsides off (and fund academia to help them) to keep game devs interested with more optimised software techniques like ReSTIR + derivatives, and performance enhancing tricks like DLSS.

maddie · May 21, 2024

Kepler_L2 said:
Unless "The Latest Thing" is superior DX12/Vulkan performance, FP16 or real Async Compute hardware. Then it doesn't matter

That brings back memories. The async thingy really awoke Nvidia's propaganda messaging for minimizing it's usefulness. Sadly, it worked well too (the propaganda).

Vattila · May 22, 2024

AMD announced yesterday that "Microsoft is using VMs powered by AMD Instinct MI300X and ROCm software to achieve leading price/performance for GPT workloads".

By the way, Microsoft Azure's November 2023 Eagle system, based on 14400 Nvidia H100 GPUs, ranks #3 on the Top500 supercomputer list. Apparently, Microsoft is now able to seamlessly plug AMD Instinct MI300X GPUs into the same system architecture. According to the Azure CTO, their AI system is now "orders of magnitude" bigger. They are currently deploying the equivalent of "5 of those supercomputers every month". Crazy times!

Page 4 - Speculation: SYCL will replace CUDA

Page 4 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

SolidQ · May 23, 2024

Antilag 2 in CS2, now waiting FSR 3.1 Comptuex maybe?

soresu · May 23, 2024

SolidQ said:
Antilag 2 in CS2, now waiting FSR 3.1 Comptuex maybe?

3.1 was announced at GDC already, 2 months ago.

Mahboi · May 23, 2024

soresu said:
3.1 was announced at GDC already, 2 months ago.

I lament this situation.
Crazy as it was to just driver hack into the games, it sounded so cool. Anti Lag + everywhere forever without any "implementation", just have it on once in driver and never care.

The Chad Anti Lag+ hack versus the Virgin Anti Lag 2 that requests implementation in a mailing list.

GodisanAtheist · May 23, 2024

Mahboi said:
I lament this situation.
Crazy as it was to just driver hack into the games, it sounded so cool. Anti Lag + everywhere forever without any "implementation", just have it on once in driver and never care.

The Chad Anti Lag+ hack versus the Virgin Anti Lag 2 that requests implementation in a mailing list.

- AMD shoulda just kept AL+ and had a couple click through warnings about using it with games that have Anti-Cheat in addition to AL2.

Would have been a cool driver level feature for older games that are never going to get a in game implementation.

Mahboi · May 23, 2024

GodisanAtheist said:
- AMD shoulda just kept AL+ and had a couple click through warnings about using it with games that have Anti-Cheat in addition to AL2.

Would have been a cool driver level feature for older games that are never going to get a in game implementation.

I also like to play Chaotic Good every time, but they'd have gotten way too many problems from idiots who wouldn't pay attention and then get VACated.

I would love a little dialogue box in Adrenalin that pops when you click AntiLag+:
"This is a driver/kernel level hack that catches your inputs early to lower your latency. It will be detected as a cheat by most online multiplayer games and you will get banned. Do you still wish to proceed?"
Then you just YOLO click it, forget it, and get banned 3 years later when you install Valorant to see what the hype about.

branch_suggestion · May 23, 2024

Lisa Su says AMD is on track to a 100x power efficiency improvement by 2027 — CEO outlines AMD’s advances during keynote at imec’s ITF World 2024

Less energy, more performance.

www.tomshardware.com

Absolute banger, will probably elaborate beyond this at Computex.

beginner99 · May 24, 2024

branch_suggestion said:
Lisa Su says AMD is on track to a 100x power efficiency improvement by 2027 — CEO outlines AMD’s advances during keynote at imec’s ITF World 2024

Less energy, more performance.

www.tomshardware.com

Absolute banger, will probably elaborate beyond this at Computex.

Form the article I get the vibe this "efficiency" increase is merely from using lower accuracy number formats like FP8

branch_suggestion · May 24, 2024

beginner99 said:
Form the article I get the vibe this "efficiency" increase is merely from using lower accuracy number formats like FP8

Only 2 of the slides are about number formats, and any future gains there are going to be very small.
The summary slide at the end says nothing about number formats being a driver of future efficiency.

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Diamond Member

Diamond Member

Member

Diamond Member

Member

Diamond Member

Member

Diamond Member

Platinum Member

Golden Member

Platinum Member

Member

Memory & Storage, Graphics Cards Mod Elite Member

Senior member

Platinum Member

Diamond Member

Senior member

Senior member

Platinum Member

Senior member

Diamond Member

Senior member

Senior member

Diamond Member

Senior member