Benchmark: MI300X multi-instance Wan 2.2 concurrency

[Stanley-blik]

Goal

Determine optimal number of concurrent Wan 2.2 instances per MI300X GPU.

Theoretical: 4-9 instances per GPU (32-72 total on 8-GPU node). Needs real hardware validation.

Tasks

• Get access to MI300X hardware (RunPod, Hot Aisle, or Azure ND MI300X v5)
• Run baseline: single instance peak VRAM + time per clip
• Scale test: 2, 3, 4 instances on same GPU — measure throughput
• Test weight sharing approach (shared model, independent working memory)
• Full node test: optimal × 8 GPUs
• Record all metrics (VRAM, bandwidth, time/clip, total throughput)
• Update docs/research/mi300x-benchmarking.md with results
• Set default max_instances_per_gpu in config based on findings

Details

See docs/research/mi300x-benchmarking.md for full protocol.

Blocked By

Access to MI300X hardware.