4x Furiosa RNGD: Qwen3-32B about 1100 tok/s aggregate at batch 32 (SDK 2026-03-13)
1100 tok/s aggregate across the batch, 4-card server
Fields
id
benchmarks/furiosa-4xrngd-qwen3-32b-b32-tps-aggregateMetric
tps_aggregate
Value
1,100
Unit
tok/s aggregate across the batch, 4-card server
Model
Qwen3-32B
Methodology
VENDOR, approximate. Same blog and same four-card RNGD server as the batch-64 cell. Qwen3-32B, SDK through 2026-03-13: batch 32 improved from approximately 750 TPS on the 2026-02-14 SDK to approximately 1100 TPS. Aggregate across the batch, not per-user. Prompt and output lengths are not stated.
Measured by
vendor
Reproducible
no
notes
Prior SDK value of about 750 TPS at batch 32 is kept in this note. Dividing 1100 by 32 gives about 34 tok/s per user if the batch is uniform, which sits inside the blog 20-40 TPS-per-user SLO band, but that division was not measured and is not the stored value.
Notes
Prior SDK value of about 750 TPS at batch 32 is kept in this note. Dividing 1100 by 32 gives about 34 tok/s per user if the batch is uniform, which sits inside the blog 20-40 TPS-per-user SLO band, but that division was not measured and is not the stored value.