“Chips from Google, Amazon, and NVIDIA historically use HBM memory and are optimized for high throughput at the expense of latency.”