Intel Arc Pro B60 24GB, llama.cpp b8175 SYCL+Vulkan: Qwen3.5-9B Q4_K_XL decode tok/s (L1 measured)

Fields

id

benchmarks/comm-arc-pro-b60-sycl-qwen35-9b-decode

Metric

decode_tok_s

Value

23

Unit

decode tok/s (tg128, SYCL backend; Vulkan 17.8)

Model

Qwen3.5-9B Q4_K_XL (GGUF)

Methodology

THIRD-PARTY MEASURED (Level1Techs, llama.cpp b8175). Qwen3.5-9B Q4_K_XL: SYCL pp2048 328.7 / tg128 23.0; Vulkan pp2048 559.6 / tg128 17.8. SYCL is better for token generation, Vulkan better for prompt processing. Qwen3.5-35B-A3B Q8: SYCL tg128 8.3 - chat-viable only for MoE. Same tester's RTX PRO 4500 did 133 t/s on the 35B for calibration.

Measured by

third_party

Reproducible

yes

notes

Best new-card $/GB in the list (~$27/GB at ~$650 in stock at Newegg). BUT ~456 GB/s means ~1/3 the decode speed of a 3090 on the same model. 24GB GDDR6, 192-bit, 456 GB/s, 200W, PCIe 5.0 x8 electrical. Multi-card: an L1 user runs 8x B60 on a PEX88096 PCIe switch + consumer X870 - works but needed the switch and Intel's BKC/platform was painful (recommends kernel 7.1+ / compute-runtime >=26.22). Buy for VRAM-bound experiments or SR-IOV/multi-instance, not for fast single-stream chat. MaxSun B60 Dual 48GB $1,939 is interesting for capacity but halves per-GPU bandwidth again. FORMAT_ID null: the source names 'Q4_K_XL' (Unsloth Dynamic XL quant); no data/quantization record exists for the UD-*_K_XL family, so format_id stays null rather than aliasing to gguf-q4-k-m, which is a different bit allocation.

Notes

Best new-card $/GB in the list (~$27/GB at ~$650 in stock at Newegg). BUT ~456 GB/s means ~1/3 the decode speed of a 3090 on the same model. 24GB GDDR6, 192-bit, 456 GB/s, 200W, PCIe 5.0 x8 electrical. Multi-card: an L1 user runs 8x B60 on a PEX88096 PCIe switch + consumer X870 - works but needed the switch and Intel's BKC/platform was painful (recommends kernel 7.1+ / compute-runtime >=26.22). Buy for VRAM-bound experiments or SR-IOV/multi-instance, not for fast single-stream chat. MaxSun B60 Dual 48GB $1,939 is interesting for capacity but halves per-GPU bandwidth again. FORMAT_ID null: the source names 'Q4_K_XL' (Unsloth Dynamic XL quant); no data/quantization record exists for the UD-*_K_XL family, so format_id stays null rather than aliasing to gguf-q4-k-m, which is a different bit allocation.

Cross-references

Intel Arc Pro B60 24GB, llama.cpp b8175 SYCL+Vulkan: Qwen3.5-9B Q4_K_XL decode tok/s (L1 measured)

Fields

Notes

Cross-references

Accelerators (1)

Engine (1)

Interconnect (1)