Compare one request
Pick a setup, a prompt and an answer. The simulator replays the measured rates: first the prompt is read, then the answer is written.
Prompt length
Answer type
Answer length (tokens)
Clock
One request, replayed
reading the promptwriting the answer
PXA 4.0PXQN4
first token 2.16 sdecode 118.0 t/stokens 512 / 512total 6.50 s
PXA 3.1PXQN4
first token 3.35 sdecode 117.7 t/stokens 512 / 512total 7.70 s
Community forkits own Q4_0 file
first token 2.52 sdecode 100.6 t/stokens 512 / 512total 7.61 s
Finish order: 1. PXA 4.0 at 6.50 s · 2. Community fork at 7.61 s · 3. PXA 3.1 at 7.70 s
| Engine | First token | Decode | Total | Time saved |
|---|---|---|---|---|
| PXA 4.0 | 2.16 s | 118.0 t/s | 6.50 s | 1.20 s |
| Community fork | 2.52 s | 100.6 t/s | 7.61 s | 0.08 s |
| PXA 3.1 | 3.35 s | 117.7 t/s | 7.70 s | — |
Simulation: replays rates measured on our box (ledger rows #8608, #8623, #8609, #8624, #8610, #8625). Not a live run. Prompt of 4,096 tokens, as measured. Answer of 512 tokens at the measured prose decode rate, held constant. At the 64k prompt, every engine uses an 8-bit KV cache.