A startup aiming to prevent programmers from being locked into using Nvidia chips has raised $6 million in funding.
London-based Spectral Compute is developing a software framework to enable applications written for Nvidia chips to run on any graphics processing unit.
Nvidia’s chips, known as GPUs, have become the dominant hardware in the AI boom. Its proprietary programming platform, CUDA, makes it possible for regular applications to use GPUs for general computing tasks — not just for graphics.
However, programs written for CUDA don’t work on other chips, such as those made by AMD or Intel. For Nvidia, CUDA is a defensive moat. For companies, switching applications written in CUDA over to non-Nvidia chips can be a lengthy and complex process.
Spectral Compute is aiming to lower the barriers for companies looking to run CUDA applications on other types of chips.
“We’re building something that’s compatible with CUDA, so software written for and tested on Nvidia’s ecosystem — their hardware — that works out of the box on competing hardware,” Michael Søndergaard, the CEO of Spectral Compute, told Business Insider.
The startup was founded in 2018 by Søndergaard, Chris Kitching, Nicholas Tomlinson, and Francois Souchay.
So far, the company’s technology only works with some AMD chip architectures. Over time, it hopes to support a wider range of hardware, including chips from Intel and semiconductor startups.
Søndergaard says its framework, called SCALE, is free for non-commercial use and that it has partnerships with research institutions for use in academia. For commercial applications, the company licenses its software based on the deployment type, working with cloud service providers, and intends to offer bespoke models for large on-premise enterprise deployments.
A separate attempt to create CUDA compatibility is ZLUDA, an open-source project that had been funded by AMD — only for the chip company to later withdraw its support and request that some code be removed.
Søndergaard told Business Insider that Spectral Compute’s approach is different.
“We’re taking a source-by-code approach,” he said. “So the original CUDA code authored by and owned by — importantly — the people who wrote that CUDA code. We’re obviously subject to licensing terms for all the code that we work with. And we don’t use Nvidia’s platform in ways that we’re not allowed to.”
Spectral Compute’s seed funding round was led by Costanoa, with participation from Crucible and prominent angel investors.
The startup plans to use the injection of capital for product development, accelerating its go-to-market expansion, and growing its 19-strong team.
Here’s an exclusive look at the 13-page pitch deck Spectral Compute used to raise its $6 million seed round.