Theoretical Bottlenecks for Scaling LLM Inference to Get Higher Token per Second

LLM inference performance is governed by three competing bottlenecks: compute time, memory bandwidth, and communication latency. In this post, we've covered what allows full hardware utilization and key constraints.