Lab notes from our B70 test queue, published as written (2026-10-15). Raw logs are not included. Back to the performance guide.
All pass lines present in RESULT.txt: guard imports branch from site-packages, Quantizing 1 layers ... lm_head,
Selected XPUW8A16FP8LinearKernel for Fp8PerChannelWeightOnlyOnlineLinearMethod, probe shows lm_head
torch.float8_e4m3fn (2048, 248320). No xe fault lines in the kernel log.
| arm | 004 | 005 | delta |
|---|---|---|---|
| B: GEMV, 4K, decode code/prose tok/s | 86.6 / 87.2 | 89.4 / 90.2 | +3.2% / +3.4% |
| B: gsm8k 100 | 97/100 | 96/100 | -1 (>= 95 bar, within noise) |
| B: model loading | 20.18 GiB | 19.71 GiB | -0.47 GiB (bf16 head freed) |
| C: GEMV + MTP k=4, 2K, decode | 157.3 / 118.7 | 183.7 / 131.9 | +16.8% / +11.1% |
| C: model loading | – | 21.75 GiB | MTP drafter’s own lm_head also quantized (“Quantizing 2 layers”) |
Note: C gains more than the head alone suggests; the drafter runs the lm_head every draft step, so its fp8 head counts k+1 times per step.