Spectral Compute

Nerds taming the green dragon with SCALE, our framework for compiling CUDA codebases for AMD GPUs, with support for more accelerated platforms coming soon.

[Spectral Compute @SpectralCom]

Cross-architecture from a single codebase is exactly why we built SCALE. Thrilled to see @AtlasInference getting this running! More performance optimizations for both @AMD and @nvidia are on the way.

https://t.co/hQpeCzwg2B

[Azeez @AtlasInference]

Atlas Inference is running Qwen3.6-27B on AMD Strix Halo 🥳

Using @SpectralCom's SCALE ROCm backend, our CUDA kernels compile and run on RDNA⚙️

Cross-architecture inference from ONE codebase 🗣️

Thank you @AIatAMD for the gift 🙏

POC ✅ excited to keep tuning performance⚡️

[Spectral Compute @SpectralCom]

@AtlasInference @AIatAMD This is just the beginning! Huge shoutout to the Atlas team for this POC. We're working hard to unlock even more performance on both @AMD and @nvidia. Let's keep pushing.

https://t.co/hQpeCzwg2B

[Spectral Compute @SpectralCom]

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.

[Spectral Compute @SpectralCom]

Hardware freedom shouldn't cost you performance. Increasingly, it buys you some.

Benchmarks here: https://t.co/kq6BZuJPGS

[Spectral Compute @SpectralCom]

New benchmark numbers are in, and SCALE keeps pulling ahead.

[Spectral Compute @SpectralCom]

And on NVIDIA's B300 (CUDA 13), SCALE lands within a whisker of nvcc on its home turf: 𝗼𝗻 𝗽𝗮𝗿 𝗼𝗻 𝗮𝘃𝗲𝗿𝗮𝗴𝗲, 𝘂𝗽 𝘁𝗼 𝟵% 𝗳𝗮𝘀𝘁𝗲𝗿 on individual workloads.
That's native CUDA tooling, matched and occasionally beaten, by a third-party compiler.

[Michael Søndergaard @SpectralMichael]

Everyone says CUDA can't target TPUs.

What they mean is nobody has written the compiler that raises CUDA code to something a systolic backend can consume.

Those are very different sentences.

Full post — Part 3 of why @SpectralCom exists: https://t.co/IAhMNaMqD4

[Spectral Compute @SpectralCom]

Slides here: https://t.co/TieAPlfikP

[Spectral Compute @SpectralCom]

@ChrisKitching17, our CTO and co-founder of Spectral Compute, recently presented at the 𝗡𝗛𝗥 𝗣𝗲𝗿𝗳𝗟𝗮𝗯 𝗦𝗲𝗺𝗶𝗻𝗮𝗿 hosted by NHR@FAU. The recording is now online: https://t.co/9rD0gDVLla

[Spectral Compute @SpectralCom]

The Q&A at the end is worth the watch on its own.

Attendees asked about:
↳ Whether optimized deep learning kernels flow through the same pipeline
↳ Support for Intel GPUs and emerging hardware
↳ How SCALE compares to NVIDIA's own compiler when targeting NVIDIA
↳ Adding custom accelerators (RISC-V came up) and emulating missing functionality
↳ Code-level transformations to improve occupancy on AMD
↳ Plans for OpenACC, CUDA Fortran, and OpenMP

[Spectral Compute @SpectralCom]

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.

[TensorWave @tensorwave]

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 - https://t.co/CIGuIPGTEy

[Spectral Compute @SpectralCom]

Our compiler toolchain is free for research and evaluation. If you maintain a CUDA project and want to finally close that thread, come say hi: https://t.co/sWKIICCvaj

[Spectral Compute @SpectralCom]

Sort the issues of almost any popular open-source CUDA project by most commented, and you'll inevitably find the exact same unresolved thread: 'Any chance of AMD support?'"

[Spectral Compute @SpectralCom]

For most maintainers, that thread stays open indefinitely. Porting a complex CUDA codebase is a part-time job, and volunteer contributors rarely have one to spare.

𝗦𝗖𝗔𝗟𝗘 𝘁𝘂𝗿𝗻𝘀 𝘁𝗵𝗮𝘁 𝗶𝘀𝘀𝘂𝗲 𝗶𝗻𝘁𝗼 𝗮 𝗿𝗲𝗰𝗼𝗺𝗽𝗶𝗹𝗲.

[Spectral Compute @SpectralCom]

𝗦𝗖𝗔𝗟𝗘 𝟭.𝟳 𝗶𝘀 𝗼𝘂𝘁.

Multi-architecture AMD builds. Kokkos support. Managed memory over 100× faster on HPC workloads. More PyTorch coverage. OpenGL interop. FFmpeg on AMD.

A lot in one release - full details in the blog post:

https://t.co/WYkVniRPWg

[Spectral Compute @SpectralCom]

A good read from our CEO @SpectralMichael.

Tell us what you think! 👇

[Michael Søndergaard @SpectralMichael]

Obvious take: if agents can write native code for any GPU, who needs a portable CUDA toolchain? Just point the model at each GPU. I think that's exactly backwards — and wrote up why. https://t.co/wzODqYZFqE

[Spectral Compute @SpectralCom]

Awesome. Love working together to enable vendor optionality for Atlas! 💻��

[Azeez @AtlasInference]

@AMD @SpectralCom we're looking forward to using SCALE for the CUDA→ROCm path :)

We want to bring this same concept of having hyper-optimized kernels built for a certain model architecture and its corresponding quantization, adapted per chip.

[Spectral Compute @SpectralCom]

That kind of structured feedback is what agents need to reason about a codebase reliably. Modern CUDA tooling is a prerequisite for agentic workflows in CUDA development.

• 𝗟𝗲𝗮𝗿𝗻 𝗺𝗼𝗿𝗲 𝗮𝗯𝗼𝘂𝘁 𝗦𝗖𝗔𝗟𝗘: https://t.co/ACBNwFnYSw

• 𝗝𝗼𝗶𝗻 𝗼𝘂𝗿 𝗗𝗶𝘀𝗰𝗼𝗿𝗱 𝗰𝗼𝗺𝗺𝘂𝗻𝗶𝘁𝘆: https://t.co/dfRZHTUhcD

[Spectral Compute @SpectralCom]

If your team is integrating AI coding agents into your CUDA development workflow this year, those agents need tooling that gives them something useful to read.

[Spectral Compute @SpectralCom]

𝗦𝗖𝗔𝗟𝗘'𝘀 𝗰𝗹𝗮𝗻𝗴𝗱 𝗶𝗻𝘁𝗲𝗴𝗿𝗮𝘁𝗶𝗼𝗻 exposes real-time semantic analysis, precise diagnostics, and cross-reference navigation on unmodified CUDA code.