Spectral Compute

A lot of people assume that going cross-vendor means settling for lowest-common-denominator performance. @SpectralMichael breaks down why that's wrong in his latest blog post:

🇸🇬 Spectral is at @superai_conf 2026 — Marina Bay Sands, Singapore, 10–11 June. Singapore is one of the fastest-moving AI hubs in the world, and the conversations here about compute capacity and GPU vendor optionality are exactly the ones we care about. If you're around, let's talk.

SCALE compiles unmodified CUDA code natively for AMD and NVIDIA. The numbers: ⚡ Up to 33.8× faster than HIP on AMD MI300X ⚡ Up to 9% faster than nvcc on NVIDIA B300 One codebase. Any GPU. No rewrites. See you in Hamburg 👋

Spectral Compute is at ISC High Performance 2026 — booth C50, Hamburg, June 22–26. 📅 Book a slot with us: scale-lang.com/s/booth-c50-ap…

COMPUTEX always delivers on serendipity: the unplanned hallway chats turn into the most valuable ones. Heading home with a full notebook and a longer list of people to follow up with. Until next time. 👋

Attendees were especially interested in the performance boost and better-than-native developer experience they can unlock with scale-lang.com - while expanding their compute capacity without being locked to a single vendor.

That's a wrap at @computex_taipei! Great few days talking about the future of heterogeneous accelerated computing.

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. scale-lang.com

@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. scale-lang.com

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 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.

Hardware freedom shouldn't cost you performance. Increasingly, it buys you some. Benchmarks here: tinyurl.com/2j8vftsr

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.

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

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: tinyurl.com/mtnxcjsw

Slides here: hpc.fau.de/files/2026/04/…

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

@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: youtu.be/uSLD40GX5nM

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.