AI BriefingB70 lags 71% behind 3090
Key point
In llama.cpp benchmarks, the Arc Pro B70 is significantly slower than the 3090, but SYCL improves things somewhat.
Details
This is a llama.cpp benchmark comparing RTX 3090 (Vulkan), Arc Pro B70 (Vulkan), and Arc Pro B70 (SYCL) on the same machine running NixOS.
On pp512, the B70 significantly lagged behind the 3090 across most models. The average shown in the table was -71.1%, though some models were faster on SYCL than on Vulkan. For example, gemma-4-E2B-it improved by +50.3%, Qwen3.5-4B by +23.5%, and Qwen3-Coder-7B by +1.4%.
Other models, however, did not turn around even on SYCL. gemma-4-26B-A4B was -16.0%, while both Qwen3.5-35B-A3B and Qwen3.6-35B-A3B were slower at -49.7%.
The first published figures for the tg128 segment show the same pattern. On Llama-2-7B, the 3090 hit 137.92 tok/s, B70 Vulkan hit 58.61 tok/s, and B70 SYCL hit 92.39 tok/s — SYCL was +57.6% faster than Vulkan, but still fell short of the 3090.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Unsloth Dynamic 3.0 GGUF Released
llama.cpp LLM Inference Accelerated 11–16x on Apple Silicon macOS VMsWhat Does GGUF Contain Besides Weights, and What's Still Missing