Reports have surfaced in the last couple of days about the concerted attempts of large Tech companies including Intel, Qualcomm and Google to break Nvidia’s dominance in data center AI chips by creating an open-source software platform on which developers can build AI models and apps.
Indeed, Nvidia’s monopoly on the booming AI training segment comes from its top-notch server-class GPUs but also from its software ecosystem, which is by far the “stickiest part” of Nvidia’s products portfolio. At the root of Nvidia’s software stack is CUDA, the company’s proprietary programming language used by millions of coders around the world to write programs that run on GPUs and hence fully exploit their superior vector and parallel computing capabilities. CUDA is used to accelerate parallel computing by breaking down a task into thousands of smaller threads executed independently.
Since its release more than 15 years ago, CUDA has become the de-facto standard for GPU acceleration in deep learning and AI applications and makes it easy for developers to take advantage of all the latest GPU architecture innovations. Accordingly, the programs and libraries relying on CUDA are countless and can be found in every piece of AI/Machine learning software.
As Nvidia’s AI supremacy is starting to be a concern for many companies, a working group, called the Unified Acceleration (UXL) Foundation, initially formed by Intel, Arm, Fujitsu, Google Cloud, Imagination Technologies, Qualcomm, and Samsung was launched in September 2023. The aim of this foundation is to further develop Intel’s oneAPI initiative, an open source and cross-platform programming model for (AI) accelerators.
This unified software development environment, based on open standards, will support heterogeneous architectures (CPU, GPU, FPGA) allowing to run its code without rewriting and/or recompiling. This “Esperanto” for (AI) accelerators is obviously at odds with CUDA which runs only on Nvidia’s chips. According to UXL’s technical steering committee, the oneAPI software platform (which has been in development for several years) will be delivered in a “mature state” by the end the year.
Overall, the strategy of UXL members, which consists in proposing an alternative to Nvidia’s software ecosystem that keeps developers tied to Nvidia chips, makes sense. That said, we believe that a migration out of Nvidia, if any, will take time as developers have a long history working on CUDA and as one of the main Nvidia rivals, AMD, is surprisingly absent from the UXL consortium (AMD is pushing its own platform called ROCm).
Nvidia has nevertheless acknowledged the threat by recently modifying the conditions of its EULA contracts (End User License Agreement) by forbidding any attempt of “reverse engineering, decompiling or disassembling any portion of the output generated using Software elements for the purpose of translating such output artifacts to target a non-NVIDIA platform”.
We then stick to our view that rival AI chips are likely to gain share over the years but that the risk for Nvidia is very manageable in the short term amid a massively expanding addressable market.
