GLM-5.3 on Two DGX Stations Peaks at 607.8 tok/s but 32 Concurrency Causes 453s Latency

petrusenko_max · x · 2026-11-15

X user petrusenkomax benchmarks GLM-5.3 on two DGX Stations: it peaks at 607.8 tok/s at concurrency 16, but at 32 concurrency throughput drops to 385.0 tok/s with 453.5-second mean latency. He argues this is effectively a 16-slot queue, not shared infra: subagents and retries serialize, and the second station is capacity you paid for but can't use.

Original post →

GLM-5.3 on Two DGX Stations Peaks at 607.8 tok/s but 32 Concurrency Causes 453s Latency

More from Infra