← Back to feed

Spectral Compute, a 30-person London compiler startup founded in 2018, has built what may be the most serious challenge yet to the CUDA lock-in that underpins Nvidia’s software moat. Its product, SCALE, is a drop-in replacement for NVCC — Nvidia’s CUDA compiler — that lets developers recompile existing CUDA code to run on AMD GPUs and, soon, other AI accelerators, without modifying the codebase.

The company claims a benchmark showing nearly 6x performance over AMD’s own HIPIFY migration tool on AMD hardware. A more aggressive number than it sounds: HIPIFY is not a foreign competitor. It is the official AMD-backed path for CUDA portability, and Spectral says it beats it by roughly 500%.

Why HIPIFY Falls Short

CUDA is not just a programming language. It exposes Parallel Thread Execution (PTX), a low-level assembly layer that unlocks deep hardware-specific optimisation. HIPIFY, AMD’s CUDA-to-HIP translation tool, largely ignores PTX, which is where many performance-critical kernels live. Intel’s SYCLomatic migrates about 90% of code, requiring manual work for the rest. ZLUDA operates on compiled binaries as a middleware layer, which imposes its own overhead. All three are translation tools. SCALE is a compiler.

The distinction matters. Spectral uses Clang and LLVM, the same compiler infrastructure that lets C++ code run natively on AMD and ARM CPUs without a performance gap attributable to anything other than the hardware itself. The clean-room implementation targets PTX, validates numerical output against NVCC, and then applies hardware-specific optimisations downstream. The founders had 60 years of combined HPC experience when they started Spectral; their background is in GPU optimisation at AI firms, not CUDA emulation wrappers.

SCALE has already run on Frontier, the exascale supercomputer at Oak Ridge National Laboratory.

The Strategic Position

What makes Spectral unusual is that it has positioned itself as neutral. It supports AMD GPUs as the primary target. It also supports Nvidia GPUs — on the thesis that Nvidia’s own compiler deliberately leaves performance on the table to sell more hardware. The company entered the Nvidia Inception program in June 2026, meaning Nvidia formally tolerated the partnership rather than opposing it. Spectral now has a commercial relationship with AMD and a programme membership with Nvidia.

CUDA’s market share is the anchor. Spectral estimates 80% of HPC code in the wild is written in CUDA. The installed base means that every new AI accelerator entering the market — whether AMD MI-series, Intel Gaudi, or third-party accelerators Spectral has not yet named — faces the same problem: the ecosystem assumes CUDA. Rewriting is not commercially viable at scale. A recompile is.

The immediate unlock is for AI researchers running established CUDA codebases who want access to AMD hardware at lower cost. The longer unlock is for AI labs and HPC centres that want hardware flexibility without betting on a single GPU vendor.

Timing

PyTorch support is due later this month. PyTorch is the dominant AI training framework, and CUDA integration is central to how it runs on GPUs. SCALE without PyTorch works for custom kernels and HPC workloads; SCALE with PyTorch becomes applicable to the mainstream AI training and inference pipeline.

Spectral has raised $6 million since founding — a small amount relative to the infrastructure problem it is targeting. The company is actively expanding from 30 employees. The GPU market’s current structure, with Nvidia at 80%+ compute market share and AI labs facing multi-year GPU delivery queues, gives the portability argument more urgency than it has had in any prior period.

Whether SCALE can maintain its performance lead as AMD’s hardware and compiler stack improve is the open question. AMD’s ROCm software ecosystem has developed rapidly over the past two years. But Spectral has eight years of head start on a compiler that operates at a lower level than its competitors, and the benchmark lead on AMD’s own hardware suggests the gap is not closing yet.