benchmark: synthetic_gpu
d: 8192
rank: 64
streams_per_gpu: 16
steps: 100000
dtype: float32
notes: "Run via torchrun. Each GPU owns independent ordered streams; only scalar metrics are aggregated."