Rail-aligned leaf switches
Difficulty 4 of 5
64 H100 in 8 servers: tensor parallel inside each server, data parallel 8 across them. The board has two cabling plans: one leaf switch per rail or one per server. Keep one plan and the spines you need. A step must take under 3.1 s, within 132,000 USD/mo.
2 GPUs · batch 64 · job 1.5 h · 2.1 USD
Goals and limits
- p50 latency
- under 3.1 s
- Monthly cost
- under 132,000 USD/mo
Concepts
How players did
Player results load from the server.