Alternative(s) to run CUDA on non-Nvidia hardware
Breaking the Chains: Spectral Compute's Quest to Decouple CUDA from Nvidia
While Nvidia is globally recognized as a hardware titan due to the massive success of its GPUs, it is equally dominant in the software realm. Its proprietary CUDA language has evolved into the de facto standard for developers working in High-Performance Computing (HPC) and Artificial Intelligence (AI).
However, a team of specialists at Spectral Compute is working to lock users into a single vendor liberate the language, enabling CUDA code to execute on non-Nvidia silicon.
The Genesis of Spectral Compute
Founded in 2018, Spectral Compute was born from the frustration of four engineers:
- Michael Søndergaard (CEO)
- Chris Kitching (CTO)
- Nicholas Tomlinson (Software Engineer)
- Francois Souchay (Software Engineer)
With a combined 60 years of experience in HPC optimization, the group was working at an AI firm when they hit a wall. They were tired of the exorbitant costs of Nvidia hardware and the lackluster performance of existing alternative compilers. This drove them to build their own solution.
The Solution: SCALE
The team developed SCALE, a tool leveraging LLVM and Clang compiler technology. It is designed to be a "drop-in" replacement for NVCC (the standard Nvidia CUDA Compiler).
While the initial focus was on AMD GPUs, the company is now expanding its reach to support various other AI accelerators. Interestingly, they still support Nvidia GPUs, arguing that Nvidia intentionally leaves performance gains on the table to incentivize more hardware sales.
The Market Landscape
According to Giulio Malitesta, Spectral's head of growth, CUDA is omnipresent.
"CUDA is basically the de-facto standard of HPC," Malitesta noted during the ISC 2026 conference in Hamburg. "We need to accept that as a fact and just do the work as compiler engineers to make it available on different platforms... but also improve on Nvidia GPUs."
Comparison of CUDA Translation Tools
Several other tools exist to migrate CUDA code, but Spectral argues they all have significant drawbacks:
| Tool | Origin | Method | Primary Weakness |
|---|---|---|---|
| HIPIFY | AMD | Translates CUDA C++ for ROCm/HIP | Ignores PTX (Nvidia assembly), limiting deep hardware support. |
| SYCLomatic | Intel | Migrates CUDA DPC++ | Only migrates of code; requires manual fixes for the rest. |
| ZLUDA | (Ex-AMD) | JIT Compiler for binaries | Operates as a middleware layer, which degrades performance. |
Technical Superiority and Performance
Spectral claims that SCALE outperforms these alternatives by utilizing a clean-room re-implementation. By applying industry-standard CPU compiler logic to GPUs, they ensure that the only performance gaps are those caused by the physical hardware differences, not the software.
The Performance Gap
The results are stark. Benchmarks indicate that SCALE provides a massive boost over HIPIFY when running CUDA code on AMD hardware:
The Workflow
The process follows a rigorous verification path:
To ensure accuracy, Spectral verifies that the numerical output of their compiled code matches the output of the original NVCC compiler.
Future Roadmap and Strategic Partnerships
Based in London and backed by $6 million in funding from last year, Spectral Compute is aggressively expanding.
Current Objectives:
- Integrate support for unnamed third-party AI accelerators.
- Deploy proprietary compiler optimizations for Nvidia GPU users.
- Release support for PyTorch (expected later this month) to better serve the ML community.
Despite their goal of breaking hardware lock-in, Spectral maintains a diplomatic relationship with the industry giants. In June, they joined the Nvidia Inception program.
"We're on the good side of Nvidia and we're on a good side with AMD," stated Ruben van Dongen, head of academic solutions and business development.
Related Industry News:
- Cloudera & VAST Data: Unified AI Factory
- Quantinuum & Rolls-Royce: Hybrid Quantum-HPC
- Intel: $5B European Manufacturing Expansion
