Intel Gaudi Optimizing Training Platform Guide

Optimization methods for Gaudi: (1) MPI CPU binding: --bind-to core --map-by socket:PE=6 (6 cores per Gaudi on HLS-1 with 96 CPUs, 8 devices). CPU affinity improves performance for both distributed and single-Gaudi. (2) Set CPU governor to performance: echo performance | sudo tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor. (3) Gaudi 3 on Sapphire/EMR Rapids: wrmsr -a 0x1b0 0x0 (Energy Perf BIAS), wrmsr -a 0x774 0x2708 (Core Frequency HWP). (4) Gaudi 3 on Granite Rapids: Disable SubNUMA in BIOS, disable C6 state, set Latency Optimized Mode, set Energy Perf BIAS to performance. (5) Ensure system components (FW, SW, BMC, CPLD) are updated. (6) Use NVMe/SSD for datasets, model code, output. This guide is referenced by vLLM Gaudi quickstart for best performance on HPU.