title: the same draft on the other engine, time per token against output on the L4
source: manuals/sglang-101/sections/4-4-speculative-decoding.md
lines: 149-154
lang: text
note: Compare the two time per token columns with the two output columns at 32 and 64.
---
| Concurrency | Default median TPOT | EAGLE-3 median TPOT | Default output tok/s | EAGLE-3 output tok/s |
|---|---|---|---|---|
| 1 | 60.16 ms | 25.61 ms | 13.42 | 31.81 |
| 8 | 66.01 ms | 35.58 ms | 112.62 | 181.55 |
| 32 | 96.18 ms | 43.82 ms | 270.11 | 269.42 |
| 64 | 107.42 ms | 51.50 ms | 250.85 | 245.34 |
