SGLang 0.5.22
NVIDIA B300 x1 · tp1 × pp1Eco-friendlyProfitable
- Speed per user (c=1)
- 252.5tok/s
- Time to first token
- 45ms
- Peak throughput
- 18,061tok/s @ 256
- Per GPU at peak
- 18,061tok/s
- Cold start
- 51s
- Context tested
- 8,192tokens
Measured 9 Nov 2026 · Throughput sweep