causal

Batching Amortizes Weight-Loading Cost

Inference is memory-bound because each parameter must be loaded from memory and used once, so batching many users to reuse each loaded weight amortizes the expensive load, and a transformer ASIC can batch around 2,500 users versus about 64 on GPUs.

causalpending

Speaker

Gavin

Evidence Quote

“You can have a much, much larger batch, not 64 but 2,500.”

Source

Gavin Uberti - Real-Time AI & The Future of AI Hardware - [Invest Like the Best, EP.356]— Invest Like the Best
Created: 11/17/2026, 12:35:53 PM

Your notes

Loading notes...