4x Furiosa RNGD: Qwen3-32B about 1500 tok/s aggregate at batch 64 (SDK 2026-03-13)
1500 tok/s aggregate across the batch, 4-card server
Fields
id
benchmarks/furiosa-4xrngd-qwen3-32b-b64-tps-aggregateMetric
tps_aggregate
Value
1,500
Unit
tok/s aggregate across the batch, 4-card server
Model
Qwen3-32B
Methodology
VENDOR, approximate, read from prose rather than a chart. Furiosa blog 2026-04-02: a server with four RNGD cards running Qwen3-32B, SDK tuned through 2026-03-13. Aggregate throughput at batch 64 improved from approximately 1200 TPS on the 2026-02-14 SDK to approximately 1500 TPS. Prompt length and output length are not stated. Aggregate tokens per second across concurrent requests, not per-user. The four-card RTX PRO 6000 comparison on the same page is chart-only and was not transcribed.
Measured by
vendor
Reproducible
no
notes
Approximate. The prior SDK value of about 1200 TPS at the same batch is kept in this note rather than as a second record, because both are read from the same paragraph. Quantization is not stated. Not a per-card number.
Notes
Approximate. The prior SDK value of about 1200 TPS at the same batch is kept in this note rather than as a second record, because both are read from the same paragraph. Quantization is not stated. Not a per-card number.