Blog

Flop Calculations for Multihead Attention Layer

The Batch Size Story (Episode 1): Exploring the GPU Memory Hierarchy

PagedAttention Deep Dive - Nagging Questions - Vol. 1

An H100 GPU Datacenter — A Simple Guide to Topology and Bandwidth

Plain‑language explanations and realistic numbers for 8×H100 nodes: NVLink/NVSwitch inside the box, InfiniBand between boxes, plus diagrams and quick checks.