FuriosaAI: Qwen3-32B on four RNGD versus four RTX PRO 6000, SDK through 2026-03-13

Vendor benchmark blog, 2026-04-02. Charts are images; only numbers written in prose were recorded. Qwen3-32B on a server of four RNGD cards, Furiosa SDK tuned through 2026-03-13 versus the SDK of 2026-02-14. Aggregate throughput approximately 1500 TPS at batch 64 (was about 1200) and approximately 1100 TPS at batch 32 (was about 750). At a 30 TPS-per-user SLO, time to first token is 1.1 to 2.1 seconds on RNGD versus 2.7 to 4.4 seconds on RTX PRO 6000. Prompt and output lengths are not stated. The four-card RTX PRO 6000 comparison is chart-only and was not transcribed.