Post by Y Combinator

Baud is building a new chip for accelerated training and inference of large AI models. Training a frontier AI model today takes thousands of GPUs, millions in capex, megawatts of electricity, and sophisticated distributed training infrastructure. This makes it such that only a handful of teams on Earth have a shot at creating frontier-level intelligence for new Opus-like moments in other domains such as robotics, drug discovery, world simulations, material science, and physics. Baud is tackling this by co-designing their chip with a new way to represent neural networks that does not require multiplications during both training and inference. This makes their chip simpler, faster, less power-hungry, and easier to build. Their first chip is validated on the Global Foundries 12nm process and is scheduled for tape-out by the end of this year. They are now taking reservations for their accelerated training service, which is already live on a cluster of FPGAs emulating their chip. Baud is founded by Sarang Z., who is a multi-time founder with 4 patents and 7+ years in building deep learning hardware, and Eric Taylor, an ex-NVIDIA chip design veteran with 4 tape-outs on leading process nodes and multiple IP releases, patents, and publications. Congrats to the team on the launch! 🚀 https://lnkd.in/gexPYcZ8

Video Content