AI BriefingSYCL Benchmark Released
·2026.04.20 01:03
Key point
On Intel Arc B70, the SYCL backend recorded pp4096 301.39 t/s and tg512 24.62 t/s.
Details
This is a llama-bench result running Intel Arc B70 on a PCIe 3.0 motherboard (H P Z640, Ubuntu 26.04 beta).
The test was conducted with the Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf model, using -d 8192, -p 4096, -n 512, --flash-attn 1 conditions.
- pp4096 @ d8192: 301.39 ± 2.92 t/s
- tg512 @ d8192: 24.62 ± 0.07 t/s
Batch settings were measured across combinations in the b=512~4096, ub=512~2048 ranges, and the currently published table contains results for the ngl 99, n_batch 512, type_k/type_v=q8_0 combination on the SYCL backend.
The author stated that they plan to subsequently release additional benchmarks with the context window increased to 131072 under the same conditions, as well as a comparison with the Vulkan backend.