Nvidia is known as a hardware company, but its CUDA language dominates AI and HPC software. Spectral Compute, founded in 2018 by engineers including CEO Michael Søndergaard, aims to detach CUDA from Nvidia and let developers run CUDA code on other chips. The team, with 60 years of HPC experience, built SCALE, a drop‑in replacement for NVCC that uses Clang/LLVM.

SCALE first targeted AMD GPUs and now expands to other AI accelerators, while also supporting Nvidia hardware. Benchmark tests show a 6x performance boost on AMD GPUs versus HIPIFY, and the clean‑room compiler approach keeps numerical accuracy intact.

The company, based in London, raised $6 million last year and employs about 30 people. It plans to add support for PyTorch, specialized CUDA libraries, and third‑party accelerators, and will release a free version for academic use.

Spectral has partnered with Nvidia’s Inception program and remains neutral, welcoming all industry players. While tools like HIPIFY, SYCLomatic, and ZLUDA exist, Spectral’s SCALE offers a more complete, performance‑friendly solution that could reshape the CUDA ecosystem.