AMD Execution Thread [2023]

[Lurkmass]

GitHub - GPUOpen-LibrariesAndSDKs/FidelityFX-SDK at release-FSR3-3.0.3

The main repository for the FidelityFX SDK. Contribute to GPUOpen-LibrariesAndSDKs/FidelityFX-SDK development by creating an account on GitHub.

github.com

Source code drop of FSR3 above in another FFX SDK branch. Vulkan is unsupported currently because it uses an obscure HLSL intrinsic that has no SPIR-V equivalent ...

 

[Qesa]

You really need to compare anything GPU with a magical power-perf-area triangle.

So does this only apply to Intel, or should we be ignoring how MI300X uses about 3x the silicon of H100 while also ignoring the single batch result?

little did Nvidia know you can also amp batch sizes on MI300X; the thing has 304 CUs after all.

 

[OlegSH]

So does this only apply to Intel

Of course, that's an AMD thread after all Conveniently ignoring RT and AI workloads is a part of that culture too)

 

[del42sa]

https://www.tomshardware.com/tech-i...try-is-motivated-to-eliminate-the-cuda-market

 

[Deleted member 2197]

https://www.tomshardware.com/tech-i...try-is-motivated-to-eliminate-the-cuda-market

Should actually go in the Intel thread ...

 

[DegustatoR]

https://www.tomshardware.com/tech-i...try-is-motivated-to-eliminate-the-cuda-market

How about eliminating the x86 market and moving to an open standard while we're at it?

 

[Bondrewd]

or should we be ignoring how MI300X uses about 3x the silicon of H100

That's how you do tiles, you spam stuff.
Also they pay more upfront costs for config flexibility which is like, the reason MI300 exists.

Of course, that's an AMD thread after all Conveniently ignoring RT and AI workloads is a part of that culture too)

someone's upset.
It's cute.

How about eliminating the x86 market and moving to an open standard while we're at it?

Replace with what exactly.
ARM is every bit as closed as x86.

 

[Shifty Geezer]

View attachment 10212

That's not really a microchip any more!

 

[Bondrewd]

Competitive performance claims and industry leading Inference performance on AMD Instinct MI300X

On December 6th, AMD launched our AMD Instinct MI300X and MI300A accelerators and introduced ROCm 6 software stack at the Advancing AI event. Since then, Nvidia published a set of benchmarks comparing the performance of H100 compared to the AMD Instinct MI300X accelerator in a select set of...

community.amd.com

and the fistfight continues

 

[Granath]

Competitive performance claims and industry leading Inference performance on AMD Instinct MI300X

On December 6th, AMD launched our AMD Instinct MI300X and MI300A accelerators and introduced ROCm 6 software stack at the Advancing AI event. Since then, Nvidia published a set of benchmarks comparing the performance of H100 compared to the AMD Instinct MI300X accelerator in a select set of...

community.amd.com

and the fistfight continues

https://twitter.com/x/status/1735742635602530629

 

[Deleted member 2197]

I wonder if AMD will supply enough information to reproduce the results. Nvidia has done so with their claim so I'd imagine some potential customers likely want to benchmark the results on their own.

Anyone wanting to validate these claims will be able to do so because NVIDIA is sharing the information necessary to reproduce the results. The blog post contains the command lines for the scripts used by NVIDIA to build its model, alongside the benchmarking scripts used to gather the data.

It’s surprising that AMD didn’t make sure that the data it shared was as accurate as necessary, as it was only a matter of time until NVIDIA or LLM enthusiasts fact checked it. If AMD wants to gain ground on NVIDIA in the race for AI market share these types of mistakes need to be eliminated.

 

[Kaotik]

That's not really a microchip any more!

No, it's several microchips

 

[Bondrewd]

https://www.amd.com/content/dam/amd/en/documents/pensando-technical-docs/article/amd-ai-networking-direction-and-strategy.pdf

Well, the holy AMD bible just dropped.
Yes, they're gonna stick their 800G IP literally everywhere and you're gonna like it.
Yes, it says xGMI over Ethernet.
Yes, xGMI memory pools.
Yes, dragonfly IF clusters. Because why not? I mean they've just hired the Cray Shasta dude.

 

[Granath]

Isn't this Tensor-RRT opensource?

 

[nAo]

Isn't this Tensor-RRT opensource?

Both TensorRT and TensorRT-LLM are open source:

GitHub - NVIDIA/TensorRT: NVIDIA® TensorRT™ is an SDK for high-performance deep learning inference on NVIDIA GPUs. This repository contains the open source components of TensorRT.

NVIDIA® TensorRT™ is an SDK for high-performance deep learning inference on NVIDIA GPUs. This repository contains the open source components of TensorRT. - NVIDIA/TensorRT

github.com

GitHub - NVIDIA/TensorRT-LLM: TensorRT-LLM provides users with an easy-to-use Python API to define Large Language Models (LLMs) and build TensorRT engines that contain state-of-the-art optimizations to perform inference efficiently on NVIDIA GPUs. Te

TensorRT-LLM provides users with an easy-to-use Python API to define Large Language Models (LLMs) and build TensorRT engines that contain state-of-the-art optimizations to perform inference efficie...

github.com

 

[Granath]

Both TensorRT and TensorRT-LLM are open source:

GitHub - NVIDIA/TensorRT: NVIDIA® TensorRT™ is an SDK for high-performance deep learning inference on NVIDIA GPUs. This repository contains the open source components of TensorRT.

NVIDIA® TensorRT™ is an SDK for high-performance deep learning inference on NVIDIA GPUs. This repository contains the open source components of TensorRT. - NVIDIA/TensorRT

github.com

GitHub - NVIDIA/TensorRT-LLM: TensorRT-LLM provides users with an easy-to-use Python API to define Large Language Models (LLMs) and build TensorRT engines that contain state-of-the-art optimizations to perform inference efficiently on NVIDIA GPUs. Te

TensorRT-LLM provides users with an easy-to-use Python API to define Large Language Models (LLMs) and build TensorRT engines that contain state-of-the-art optimizations to perform inference efficie...

github.com

Any change to HIPpyfy it ?

 

[del42sa]

AMD strikes back at Nvidia with new MI300X benchmarks — MI300X shows 30% higher performance than H100

 

[Lurkmass]

Any change to HIPpyfy it ?

They already have a TensorRT equivalent in the form of MIGraphX ...

 

[Deleted member 2197]

AMD strikes back at Nvidia with new MI300X benchmarks — MI300X shows 30% higher performance than H100

Has AMD provided any information so people can test the resulting benchmarks? That is basically the big difference between current IHV provided benchmark results.

 

[Bondrewd]

Has AMD provided any information so people can test the resulting benchmarks?

You can't until MI300X is GA at Azure and Oracle, which is next Q.