Phoronix conducted a performance comparison between Vulkan and AMD ROCm back-ends for local AI workloads. The tests utilized the Lemonade 2026.39.1 local AI server, which incorporates Llama.cpp build b10825 to facilitate benchmarking.
The evaluation involved two distinct hardware configurations. The first setup featured a System76 Thelio Major workstation equipped with an AMD Radeon AI PRO R9700 GPU. The second system was a Framework Desktop powered by the AMD Ryzen AI Max+ 395 Strix Halo processor.
Previous testing conducted nearly a year ago on AMD RDNA4 hardware indicated that the Llama.cpp Vulkan back-end outperformed the AMD ROCm back-end in several models. The current benchmarks aim to provide an updated assessment of the performance gap between these two interfaces.
