GoKawiilNvidia is primarily known as a hardware company thanks to the wild success of its GPUs. But Nvidia is also a force in the world of software, thanks to its CUDA programming language that has become the defacto standard for AI and HPC developers. Now a group of CUDA experts at Spectral Compute are looking to unhook the language from Nvidia and allow users to run their CUDA code on other chips.
Spectral Compute was formed in 2018 when a group of four engineers, including CEO Michael Søndergaard, CTO Chris Kitching, Software Engineer Nicholas Tomlinson, and Software Engineer Francois Souchay, decided they had had enough with hardware lock-in with their CUDA code. The founders, who had 60 years of combined HPC optimization experience, were working at an AI firm, but grew so frustrated with the cost of Nvidia GPUs and the poor performance of alternative compilers that they decided to build their own.
The founders developed a product called SCALE, which uses CLang and LLVM compiler technology, to function as a drop-in replacement for NVCC, the Nvidia CUDA Compiler. The company’s first target with SCALE were AMD GPUs, but now it’s broadening its goals to enable users to run CUDA on other AI accelerators. The company also supports Nvidia GPUs, on the theory that Nvidia is leaving software-based performance optimization on the table because it wants to sell more hardware.
Spectral is a big fan of CUDA, which it claims represents about 80% of the HPC code in the wild, according to Giulio Malitesta, head of growth for Spectral. “CUDA is basically the de-facto standard of HPC,” Malitesta told HPCwire at the ISC 2026 conference in Hamburg, Germany. “We need to accept that as a fact and just do the work as compiler engineers to make it available on different platforms that are not necessarily Nvidia, but also improve on Nvidia GPUs.”
There are other compilers on the market that can make CUDA code run elsewhere, as Malitesta pointed out. AMD created a tool called HIPIFY to translate Nvidia CUDA code into C++ code that can be run on AMD’s ROCm software stack via HIP (Heterogeneous-compute Interface for Portability). Then there’s SYCLomatic, a open source tool originally developed by Intel to migrate CUDA code to Data Parallel C++ (DPC++). And who could forget ZLUDA, which is a just-in-time compiler once backed by AMD that takes CUDA binaries and runs them on non-Nvidia hardware.
However, most tools have a downside. For instance, HIPIFY–which the Spectral founders worked with previously–essentially ignores Parallel Thread Execution (PTX), the Nvidia assembly language that opens up deep hardware support in CUDA, Malitesta says. SYCLomatic migrates about 90% of code, requiring manual work for the final 10%. And ZLUDA operates on the compiled binary code and acts as a middleware layer, which hurts performance. Legal questions have also dogged various non-Nvidia CUDA compilers.
Malitesta said Spectral Compute is able to surpass the middling performance of these other compilers with SCALE and enable CUDA to take full advantage of the underlying hardware. The company has published benchmarks on its website that show SCALE delivers nearly a 6x performance boost on AMD GPUs compared to using HIPIFY to convert CUDA code to AMD’s own ROCm environment.
Spectral is able to achieve such performance levels because it’s taken a clean-room re-implementation based on a cutting-edge compiler framework, Malitesta said. “We take the approach that’s industry-standard for CPUs, but apply it to GPUs,” he said, adding that it’s “the same approach that enables C++ to run, for example, on AMD and ARM CPUs, where nobody expects a performance gap that isn’t directly caused by differences in the underlying hardware.”
After recompiling the code, Spectral verifies that it’s correct from a numerical perspective. If it matches the normal NVCC output, then the company considers it a successful implementation.
Spectral, which is located in London and raised $6 million last year, is currently working on supporting third-party AI accelerators, which have not yet been named. The company is also working on some proprietary novel compiler optimizations that it says will deliver a major upgrade for customers running CUDA on Nvidia GPUs. Later this month, it’s anticipating releasing support for PyTorch, which will enable its tools to work better with the AI and machine learning framework.
... continue reading