Enthusiast benchmarks 4x Intel Arc Pro B65: full long-context throughput sweeps for Qwen3.8 models
SamSausages · reddit · 2026-10-20
With hardly any benchmarks available for the Intel Arc Pro B65, one user built a 4x B65 rig and ran systematic long-context throughput sweeps with vLLM.
- Tested Qwen3.8-Flash-Next-W4A16 and Qwen3.8-27B (INT4/FP8) across 0–200K contexts at C1–C16 concurrency, logging total and median per-request throughput
- Multi-card scaling pays off: 27B INT4 on 4 cards hits 468 tok/s aggregate at C8 with no context, but per-request throughput degrades sharply as context and concurrency grow, with queuing past capacity
- Early layer-offload-to-RAM tests (Flash-Next only so far) show interesting results
- Full matrices are public at b65.thekaiser.us; the author takes requests for specific tests
More from Infra
- Intel CTO: chip output can at best 2X while AI compute grows 3X per year — SammaVaco · 2026-10-23
- Engineer reverse-engineers projector TV's 10s wake lag to a cold-booting Linux board — banteg · 2026-10-23
- Ollama CEO says enterprise AI is having its PC moment on This Week in AI — ollama · 2026-10-23
- Modal open-sources Spindle, a Tinker-compatible training server that charges GPU time instead of tokens — josh_wills · 2026-10-23
- Aleph Alpha documents RL sandbox infrastructure lessons: stalled GPUs, full disks, crash-as-wrong-answer bugs — _ScottCondron · 2026-10-23
- Treat proprietary silicon as strategic lock-in, architect warns — DavidLinthicum · 2026-10-23