Spectral Compute adapts CUDA for AMD accelerators More Effectively than even AMD
12:15, 15.07.2026
Startup Spectral Compute, based in London, has designed a CUDA compiler that outperforms tools developed by AMD on AMD hardware.
Insights regarding the SCALE compiler
SCALE compiler functions by allowing existing CUDA code to function on other hardware, and there is no necessity for source code changes. The new compiler serves CUDA code directly to the target hardware, and there is no need to change CUDA into other languages. That means variety in the performance characteristics depends only on the hardware, but not on the compiler.
Currently, CUDA dominates in the HPC sector. However, most of the previous solutions were just converting the code. For instance, AMD HIPIFY functions by converting CUDA code to C++ for the ROCm platform. Intel SYCLomatic converts 90%, but the rest is done manually. That’s why such a new method excludes the necessity of recompiling and, in such a way, improves performance.
Now, the startup is mainly focused on the AMD hardware, but is also working with other accelerator architectures, including Nvidia’s hardware. Spectral is also working on the support for some CUDA libraries.
Hope you found this article helpful - what do you think? Like and subscribe to our blog to get more practical tips and the latest tech news from HostZealot.