01. AI Hardware & Chips

Chapter focus: The "common language" of the compute foundation. Every discussion in a cluster — memory, network speed, energy efficiency — ultimately lands on chip and board specifications. This chapter builds the analysis framework around compute / memory / interconnect, then gives engineering methods from selection to acceptance.

Articles

1. AI Accelerator Overview: Architecture, Generations & Selection — The evolution logic of the three pillars (Tensor Core / memory bandwidth / NVLink), with a selection matrix for training, inference, and scientific computing.
2. 10K-GPU Cluster Hardware Selection & Acceptance Checklist — Five-dimension selection (compute/memory/interconnect/storage/power), and a four-step acceptance method: HPL / Stream / bandwidth / 72h stability.
3. GPU Drivers & Lifecycle Management — Version matrix, gray upgrades and rollback, Fabric Manager matching, Xid/ECC monitoring and hardware lifecycle.
4. Liquid-Cooled Data Center Design — From air-cooling limits to cold-plate cooling: selection, key components, acceptance and operations.

📌 Planned: GPU driver deep-dive, PCIe/NVLink topology planning.

Reading Path

• Procurement / architecture decisions: start with the accelerator overview to build the selection matrix, then walk the acceptance checklist.
• Cluster delivery / operations: go straight to the acceptance checklist and check item by item.
• Deep architecture interest: start with the "three pillars" framework, then combine with the training chapter to understand how compute gets scheduled.

Key Takeaways

• When evaluating an accelerator, always look at three things: compute (TFLOPS), memory bandwidth (TB/s), interconnect bandwidth (GB/s) — together they determine training throughput; none can be missing.
• For training, the real lever of GPU generational upgrades is memory bandwidth: H100's 3.35 TB/s vs A100's 2.0 TB/s is a +67% jump — a bigger lever than raw compute density.
• The final test of selection is measurement: every paper spec must pass HPL (compute), Stream (memory bandwidth), and nccl-tests (interconnect), then 72 hours of stability.