Arc Pro B65 report isolates a long-prefill regression to Intel runtime packages
One community report holds llama.cpp a956a852 and Qwen3-Coder-30B-A3B-Instruct Q4_K_M constant while changing Intel packages. Long prefill falls while decode rises. No independent reproduction supports generalizing to other GPUs.
Source: llama.cpp社区单份自报
A community issue filed on November 16 reports a software-stack regression on one Intel Arc Pro B65. With llama.cpp source a956a852, model and flags held constant, replacing Intel runtime packages reduced 32K-token prompt processing while improving decode. This is neither a new GPU launch nor an AraoAI measurement. At collection time it was a single report with no independent reproduction in the comments.
The setup uses one B65, the xe driver, and Qwen3-Coder-30B-A3B-Instruct Q4_K_M. The command is llama-bench -fa 1 -b 4096 -ub 2048 -p 32768 -n 128 -r 3 with GGML_SYCL_ENABLE_XMX_ESIMD=0. The current stack includes oneDNN 2026.1.0, IGC 2.42.6 and NEO 26.40. Its table reports approximately 872–895 tok/s for pp32768 and approximately 100 for tg128. With older IGC 2.34.4 and NEO 26.18 packages while keeping oneDNN 2026.1.0, the corresponding entries are 1079 and 92. We preserve approximate values and ranges rather than manufacture precise midpoint measurements.
The useful evidence is the controlled comparison, not a GPU ranking. Reverting the initially suspected source change did not restore performance. Reverting only oneDNN did not do so either. Package dependencies prevent freely mixing the IGC and NEO versions, so the report cannot yet identify one component as the sole cause.
Prefill and decode can move in opposite directions. A claim that an upgrade is faster based only on generation speed misses long-prompt waiting time. The author explicitly leaves other GPUs, dense models and Windows unverified; the report is not a blanket recommendation to downgrade.
AraoAI is holding this as a candidate rather than importing a formal benchmark record. The original text lacks an explicit operating-system version, and several results are approximate ranges. Missing evidence should remain visible instead of being filled by inference.
Source: https://github.com/ggml-org/llama.cpp/issues/32399