←
Michael Søndergaard52 posts
Michael Søndergaard avatar

Michael Søndergaard

@SpectralMichael

CEO @SpectralCom. SCALE: your CUDA code, native on AMD and NVIDIA. The future of accelerated computing is multi-vendor; now the software is too.

◷ Joined July 2024
Michael Søndergaard avatar
Michael Søndergaard@SpectralMichael·Nov 6

HOOMD-blue has an open PR titled "Remove HIP". The maintainers are moving to plain CUDA, and for AMD GPUs the PR's first option is building it with SCALE. One CUDA codebase for both vendors is the whole point. github.com/glotzerlab/hoomd-…

Spectral Compute avatar
⟳ Michael Søndergaard retweeted
Spectral Compute@SpectralCom·Oct 7

AMD Instinct™ GPUs are incredible. A single AMD Instinct™ MI355X sports 𝟮𝟴𝟴𝗚𝗕 𝗼𝗳 𝗛𝗕𝗠𝟯𝗘, 𝟴𝗧𝗕/𝘀 𝗼𝗳 𝗽𝗲𝗮𝗸 𝗺𝗲𝗺𝗼𝗿𝘆 𝗯𝗮𝗻𝗱𝘄𝗶𝗱𝘁𝗵 𝗮𝗻𝗱 𝟳𝟴.𝟲 𝗧𝗙𝗟𝗢𝗣𝘀 𝗼𝗳 𝗽𝗲𝗮𝗸 𝗙𝗣𝟲𝟰. Put eight in a node and you have 2.3TB of HBM3E: enough to hold an entire frontier-scale model in memory on one machine. For accelerated HPC, it means bigger simulations resident on a single GPU, bandwidth-bound kernels with space to breathe, and serious double precision. SCALE compiles supported CUDA code to run natively on it, cutting the effort required to move existing applications. Our compiler is built to use what the hardware offers. One example: AMD GPUs can pass data between threads through DPP, hard-wired lane permutations that cost almost nothing. SCALE recognises the patterns in ordinary CUDA shuffles and maps them onto DPP automatically, so your code takes the fast path without a line of AMD-specific code. The silicon was always capable. What held CUDA workloads back on AMD was getting the code there without leaving performance behind on the way. That is a compiler problem, and compilers can be fixed once for everyone. We are now listed in @AMD's Instinct AI & HPC Software Ecosystem Program! amd.com/en/ecosystem/isv/dat…

Post attachment
HPCwire avatar
⟳ Michael Søndergaard retweeted
HPCwire@HPCwire·Jul 22

A group of #CUDA experts at Spectral Compute are looking to unhook the language from #Nvidia and allow users to run their CUDA code on other chips. ow.ly/nWPz50ZnBV3

Post attachment
Michael Søndergaard avatar
Michael Søndergaard@SpectralMichael·Jun 12

• GROMACS: Unmodified CUDA matches native HIP on AMD. • Rodinia: Average speed-up over HIP of 6.19x (peaking at 33.8x). • NVIDIA: On par with nvcc (peaking 9% faster).

Michael Søndergaard avatar
Michael Søndergaard@SpectralMichael·Jun 12

You can't have portability AND peak performance. That's the assumption people make. They're right if you rely on portability by abstraction at runtime. But that assumes a cross-vendor system must be a "lowest-common-denominator" system.

Spectral Compute avatar
⟳ Michael Søndergaard retweeted
Spectral Compute@SpectralCom·Jun 3

𝗦𝗽𝗲𝗰𝘁𝗿𝗮𝗹 𝗖𝗼𝗺𝗽𝘂𝘁𝗲 𝗶𝘀 𝗻𝗼𝘄 𝗽𝗮𝗿𝘁 𝗼𝗳 𝘁𝗵𝗲 𝗡𝗩𝗜𝗗𝗜𝗔 𝗜𝗻𝗰𝗲𝗽𝘁𝗶𝗼𝗻 𝗽𝗿𝗼𝗴𝗿𝗮𝗺. 🟩 Inception is @nvidia's program for AI startups - a membership that gives access to technical resources, preferred pricing on NVIDIA hardware and software, and exposure to a global network of investors and partners. CUDA is the de-facto standard for AI developers, and we’re honored to play our part in growing the ecosystem.

Post attachment
Spectral Compute avatar
⟳ Michael Søndergaard retweeted
Spectral Compute@SpectralCom·May 22

𝗨𝗽 𝘁𝗼 𝟮𝟱.𝟳× 𝗳𝗮𝘀𝘁𝗲𝗿. Unmodified CUDA. AMD silicon. Thanks to the @tensorwave team for benchmarking SCALE on MI355X and publishing the numbers. Port to AMD used to mean a rewrite. Now it means a recompile.

Spectral Compute (@SpectralCom) used TensorWave’s AMD-native infrastructure to benchmark CUDA portability and performance on @AMD Instinct™ MI355X GPUs. See how they did it - tensorwave.com/blog/spectral…

Post attachment
Michael Søndergaard avatar
Michael Søndergaard@SpectralMichael·May 13

CUDA isn't a standard. No consortium, no ratified spec. It's whatever NVIDIA ships next. That's supposed to be a weakness. It's why CUDA wins. The cross-vendor layer can't be a consortium either. It has to be a company. That's what we're building at @SpectralCom. SCALE compiles unmodified CUDA on AMD and NVIDIA today. No PDF to argue about — a toolchain. Full piece: [bit.ly/3PBsYsU](bit.ly/3PBsYsU)