The Beyond CUDA Summit, hosted by TensorWave, is underway, focusing on AI workloads utilizing AMD’s MI Instinct accelerators. This event tackles the concept of the so-called ‘CUDA moat’ and aims to provide developers with insights on optimizing their AI-centric workloads with alternatives. Notable speakers include distinguished computer architects Jim Keller and Raja Koduri, who will offer panels, demonstrations, and expert opinions throughout the summit.
Industry Context and Challenges
Nvidia-controlled GPUs dominate the AI hardware landscape, bolstered by a robust and mature CUDA ecosystem that many organizations rely on, even as AMD’s Instinct accelerators perform comparably. Nvidia capitalized on the potential of parallel computing early, creating CUDA, which has since become the standard for GPU-accelerated applications. Currently, about 90% of Nvidia’s revenue comes from its data-center business, where CUDA is a key component, resulting in a vendor lock-in that stifles innovation and competition.
Alternatives to CUDA
While the industry is moving towards a more open-source and hardware-agnostic future, challenges remain. Alternatives such as OpenCL, ROCm, oneAPI, and Vulkan lag behind Nvidia’s solutions in various areas. The Beyond CUDA Summit aims to address these challenges, discussing issues like hardware flexibility, cost efficiency, and viable alternatives to CUDA.
Specific Challenges with AMD’s Solutions
For example, ROCm faces significant hurdles before it can match CUDA’s capabilities. Currently, ROCm supports a limited selection of modern GPUs, whereas CUDA is compatible with hardware dating back to 2006. Notably, AMD’s latest RDNA 4 GPUs are still not officially supported by ROCm, and developers have expressed frustration over the slow addition of new features and hardware support by AMD. However, there is a positive note with the recent addition of Strix Halo, which is now ROCm-compatible, albeit only on Windows.
Event Details
Located at The Guildhouse in San Jose, the summit coincides with Nvidia’s GTC, occurring just three blocks away at the McEnery Convention Center. The event runs from 12 PM to 10 PM PDT, featuring multiple sessions, and attendees have the chance to win an AMD Instinct MI210 GPU equipped with 64GB of HBM2e memory. Further details about the summit can be found on the official event page.