Deep Dive into TPU and GPU Cluster Communication
gordic_aleksa · x · 2026-07-15
This extensive article explores the collective communication primitives within TPU and GPU clusters and how they support large model training and inference scaling.
Key topics covered include:
- TPU cluster topology: superpods, slices, DCN, PCIe, ICI
- Core communication operators: All-Gather, Reduce-Scatter, All-Reduce, All-to-All
- The All-to-All dependency for token distribution in MoE
- GPU cluster topology: DGX reference architecture, nodes, fat trees, scalable units
- Intra- and inter-node communication algorithms: rings, trees, SHARP, hierarchical collective communication, InfiniBand
The author emphasizes that truly understanding scaling for training/inference requires looking beyond FSDP, expert parallelism, or data parallelism to grasp the underlying hardware topology and communication algorithms.
Related event: Deep Dive on Collective Communication in TPU/GPU Clusters Gains Traction(5 posts)→
More from Infra
- SF Compute founder: buying compute is 'an absolutely awful experience' right now — IgorCarron · 2026-09-11
- SmolVM open-sources persistent computer infrastructure for agents that outlive chat sessions — aniketmaurya · 2026-09-11
- PyTorch Day Korea 2026 launches first offline conf, CFP closes Sept 13 — PyTorch · 2026-09-11
- Local LLM server dilemma: 4x CMP-170HX (price up 53% in 20 days) vs Mac Studio M5 Ultra — rumboll · 2026-09-11
- llama.cpp lands Flash Attention tuning for RDNA4, big prefill gains on AMD — pmttyji · 2026-09-11
- Your p99 latency benchmark may be lying: a deep dive into coordinated omission — Franc0Fernand0 · 2026-09-11