Recently there were Phoronix benchmarks looking at [1]the Intel Battlemage GPU compute performance since last year when the Arc B580 graphics card launched as well as [2]the OpenGL and Vulkan graphics performance for the B580 on Linux since launch. There was much progress on the open-source Intel Linux graphics drivers at large this year but especially for Battlemage. Following that a [3]Phoronix Premium reader asked about seeing some fresh Llama.cpp AI benchmarks with its Vulkan back-end now for the Arc B580 compared to competing AMD and NVIDIA graphics cards. Here are those benchmarks as requested.

[4]

Today's benchmarking is looking at the Llama.cpp Vulkan performance for the Intel Arc Graphics B580 up against the AMD Radeon RX 9000 series and NVIDIA GeForce RTX 50 graphics cards. That's based upon the graphics cards I have available (i.e. no RTX 5060 Ti or RTX 5070 Ti in my possession and keeping it to the consumer GPUs for lacking the Arc Pro B60 and RTX PRO Blackwell graphics cards). All of the latest Linux graphics drivers from each vendor were used for this Battlemage vs. RDNA4 vs. Blackwell comparison on Linux with Llama.cpp using Vulkan everywhere.

For the AMD and Intel graphics cards the latest drivers meant using Linux 6.18 and Mesa 26.0-dev for the latest kernel drivers as well as the newest Intel ANV and AMD RADV Vulkan drivers. On the NVIDIA side was using the NVIDIA 580.95.05 driver stack with its open kernel driver. Aside from the driver and GPU differences the same system was used for this fresh round of benchmarking of Vulkan for AI inferencing.



[1] https://www.phoronix.com/review/intel-b580-compute-one-year

[2] https://www.phoronix.com/review/intel-b580-opengl-vulkan-eoy2025

[3] https://www.phoronix.com/phoronix-premium

[4] https://www.phoronix.com/image-viewer.php?id=llama-cpp-vulkan-eoy2025&image=llama_vulkan_eoy2025_lrg