Copyright SiliconANGLE News
A small British software startup called Spectral Compute Ltd. believes it has what it takes to break Nvidia Corp.’s stranglehold on artificial intelligence data centers after raising $6 million in seed funding today. Spectral Compute isn’t trying to make advanced AI processors that can compete with Nvidia’s high-powered graphics processing units. Rather, its intention is to make it easier for companies to run their AI applications on alternative hardware, starting with Advanced Micro Devices Inc.’s GPUs. The problem it’s trying to solve is that most AI applications today are based on Nvidia’s Compute Unified Device Architecture or CUDA programming model, which is what makes it possible for GPUs to be used for general-purpose computing rather than for graphics processing. CUDA is essential for AI developers to properly harness GPUs to power their AI models, but the issue is that it’s owned by Nvidia. The reliance on CUDA means that most AI applications are not compatible with non-Nvidia GPUs and processors, resulting in what some have termed the “Nvidia tax.” Developers feel as if they’re forced to go with Nvidia’s GPUs, and that means paying premium costs to purchase or rent the hardware, and struggling with significant delays. Nvidia’s chips are in such huge demand that most customers trying to order them face waiting times of up to 18 months. In the world of central processing units, application portability was solved decades ago with compilers such as GCC and LLVM, which means software can easily be ported to run on chips by AMD, Intel Corp. or some other provider without any hassles. But GPUs lack such an easy solution, which means developers are forced to rewrite their application code virtually from scratch if they want to leverage hardware from one of Nvidia’s rivals. Run CUDA anywhere Or at least, that used to be the case, but Spectral Compute says it has finally come up with an alternative. Its solution is SCALE, a GPU programming toolkit and compiler that will ultimately enable CUDA-based code to run natively on any AI-accelerated chip. For now, SCALE’s compatibility is limited to AMD’s GPUs, but the company aims to expand its support to all AI accelerators in future. SCALE makes it possible for programs written in the CUDA language to be directly compiled for AMD’s GPUs in seconds, so developers can simply deploy their code as it is. While other tools do exist for porting CUDA-based code to third-party GPUs, they still require extensive changes to the underlying codebase and can take months to implement, with the risks of bugs and degraded performance. In contrast, Spectral Compute claims that SCALE enables almost instantaneous interoperability. In other words, companies will no longer be locked in to Nvidia’s chips, and can finally implement multivendor strategies and procure whatever AI hardware makes sense for them, based on performance, efficiency and availability. Spectral Compute co-founder and Chief Executive Michael Søndergaard said companies shouldn’t be forced to rely on a single chipmaker. “The vision of SCALE is that teams can finally use their existing CUDA code on any GPU, whether that’s Nvidia, AMD, Intel, or others, without performance loss or costly rewrites,” he explained. “We’re giving developers and enterprises the freedom they’ve never had before.” The round was led by Costanoa and saw participation from Crucible and a number of angel investors. The funds from today’s round will be used to accelerate the startup’s go-to-market strategy and product development, so it can bring instant portability to other AI processors. Costanoa partner Tony Liu said Spectral Compute has solved one of the toughest challenges faced by AI developers. “It lets teams write CUDA once and run it anywhere,” he said. “Its breakthrough technology eliminates one of the largest bottlenecks in AI infrastructure and HPC today, unlocking choice and innovation across the compute landscape.” Image: SiliconANGLE/Microsoft Designer
