4x Furiosa RNGD: Qwen3-32B TTFT 1.1-2.1 s at a 30 tok/s per user SLO
1600 ms, midpoint of a 1100-2100 ms vendor range, 4-card server
Fields
id
benchmarks/furiosa-4xrngd-qwen3-32b-ttft-ms-slo30Metric
ttft_ms
Value
1,600
Unit
ms, midpoint of a 1100-2100 ms vendor range, 4-card server
Model
Qwen3-32B
Methodology
VENDOR RANGE. Furiosa blog 2026-04-02: at a 30 TPS-per-user SLO on Qwen3-32B, the first token on RNGD arrives in 1.1 to 2.1 seconds, against 2.7 to 4.4 seconds on RTX PRO 6000 in the same comparison. The stored value is the midpoint, 1600 ms; the range is the claim. The blog says the half-TTFT advantage holds from batch 8 through batch 64, so this is not a batch-1 TTFT. Prompt length is not stated. Four-card server, same setup as the aggregate TPS records.
Measured by
vendor
Reproducible
no
notes
Do not read 1600 ms as a measured point. The RTX PRO 6000 side of the same sentence (2.7 to 4.4 s) is a different accelerator and does not get its own record, because card count, batch, and prompt length for that side are only in the chart. Draft because of the range and the missing prompt length.
Notes
Do not read 1600 ms as a measured point. The RTX PRO 6000 side of the same sentence (2.7 to 4.4 s) is a different accelerator and does not get its own record, because card count, batch, and prompt length for that side are only in the chart. Draft because of the range and the missing prompt length.