“We spend enormous amounts in R&D dollars to bring these chips to market. It’s hard, because even if you know how to, building a GPU costs tens of billions of dollars. But more importantly, we pursue many different chip designs, and only a very select few become production-grade. This process takes years, because AI algorithms are changing rapidly, which changes what kind of processors are ideal. Even if you’re building 10 different chips, you’ll still miss out on what’s needed. Nvidia, for instance, is building GPUs, LPUs, CPUs, and even DPUs (data processing units)—which are all varying types of silicons,” he said.
Leave a Reply