Spectral Compute (@SpectralCom) / X

Spectral Compute

340 posts

Spectral Compute

@SpectralCom

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

483 Green Lanes London England

Joined September 2025

Spectral Compute
@SpectralCom
Jul 30
𝗧𝗵𝗲 𝗲𝘅𝗽𝗲𝗻𝘀𝗶𝘃𝗲 𝗚𝗣𝗨 𝗶𝘀𝗻'𝘁 𝘁𝗵𝗲 𝗼𝗻𝗲 𝘆𝗼𝘂'𝗿𝗲 𝘂𝘀𝗶𝗻𝗴. 𝗜𝘁'𝘀 𝘁𝗵𝗲 𝗼𝗻𝗲 𝘆𝗼𝘂'𝗿𝗲 𝗵𝗼𝗹𝗱𝗶𝗻𝗴 𝗷𝘂𝘀𝘁 𝗶𝗻 𝗰𝗮𝘀𝗲. Teams reserve far more capacity than they're running, because being caught without it later costs more than paying for it
Spectral Compute
@SpectralCom
Jul 29
"Could we run this on other hardware?" is usually perceived as an entire project, in accelerated HPC. Scoping, a port plan, a spike nobody has time for. It doesn't have to start that way. Install SCALE. Point your existing build at it instead of nvcc. Run your test suite.
Spectral Compute
@SpectralCom
Jul 27
𝗔𝘀𝗸 𝗮 𝗰𝗼𝗱𝗶𝗻𝗴 𝗮𝗴𝗲𝗻𝘁 𝗳𝗼𝗿 𝗮 𝗚𝗣𝗨 𝗸𝗲𝗿𝗻𝗲𝗹 𝗮𝗻𝗱 𝘆𝗼𝘂 𝗴𝗲𝘁 𝗖𝗨𝗗𝗔 𝗰𝗼𝗱𝗲. Because that's what it read: twenty years of tutorials, forum answers, textbooks and public repos. Where there's less public code to learn from, the model guesses more, and
Spectral Compute
@SpectralCom
Jul 23
Hardware improves once per generation. A compiler improves every release. SCALE treats GPU performance as a compiler problem - and compilers improve with every release. Each novel optimization we bake into SCALE lands on every target at once: the accelerator you're evaluating
Spectral Compute
@SpectralCom
Jul 20
𝗦𝗖𝗔𝗟𝗘 𝟭.𝟳.𝟮 𝗶𝘀 𝗼𝘂𝘁. Another big step toward full nvcc drop-in compatibility - undocumented quirks included. Much faster atomic arithmetic on AMD RDNA GPUs. A new leak-detection feature. Graph Memory Nodes and a long list of new CUDA APIs. Some of the runtime's
See @SpectralCom's full profile
Sign up
Log in

See @SpectralCom's full profile

See @SpectralCom's full profile