Investors are eagerly awaiting to see how Cerebras Systems, a startup rivaling NVIDIA in the realm of AI chips, will perform in its upcoming initial public offering in late 2024. Led by Lior Susan, Eclipse Ventures’ founder and managing partner, Cerebras has developed specialized chips perfect for generative AI tasks like model training and inference, competing against NVIDIA’s B100 and H100 chips.
Nevertheless, CUDA, a software layer developed by NVIDIA to assist developers in managing their graphics processing units, poses a significant challenge to competitors like Cerebras. CUDA boasts millions of lines of code that have become ingrained within the AI ecosystem. While Cerebras has established its own software compatible with its chips, it still lags behind CUDA in terms of adoption and functionality, hampering any attempts to dethrone NVIDIA’s dominance.
This reliance on CUDA has made it difficult for developers to seamlessly transition to alternative AI chips, even if startups offer tools like Hypify to facilitate code translation. A recent attempt by a coalition of NVIDIA competitors to create a chip-agnostic software platform as a CUDA alternative highlights the industry’s struggles in breaking free from NVIDIA’s stronghold.
Despite these challenges, companies like Vimaan, focused on AI-driven warehouse automation, continue to build their technology on CUDA due to its established track record and vast developer support. While there are calls for customers to diversify their GPU sources, NVIDIA’s market share remains formidable, with long-term prospects shaping the future landscape of AI chip market.
Overall, the battle to challenge NVIDIA’s dominance in the AI chip market remains an uphill climb, as competitors face the daunting task of breaking free from CUDA’s tight grip on developers’ preferences and industry standards.
Article Source
https://www.businessinsider.com/nvidia-cuda-software-gpu-ai-chips-cerebras-ipo-2024-6