The Compute and Bandwidth Logic Behind Inference Optimization

Abhishekcur · x · 2026-07-08

This post explains why inference is memory-bandwidth bound from the perspective of arithmetic intensity. It points out two core optimization strategies: reducing the bytes transferred per token, or increasing the computation carried by each byte. Methods mentioned include MoE, batching, MLA, quantization, GQA, paged cache, speculative decoding, and expert parallelism.

Related event: How DeepSeek-R1 Achieves Faster, Cheaper Inference(3 posts)→

Original post →

More from Infra

Infra channel →