llama.cpp
76 GPUs Benchmarked: Memory Bus Width Outweighs Newer Silicon for Local AI
76 GPUs tested against second-hand prices, generation speed, and VRAM capacity. Wide memory buses beat newer cards on local LLM value.
Back to home
Vulkan
4 articles tagged with this topic
AMD
AMD GPUs Now 30% Faster on Local LLMs — Another Reason to Escape NVIDIA
llama.cpp's new Vulkan backend for AMD RDNA3/RDNA4 GPUs delivers ~27% throughput boost on 26B models. A signal for SMEs eyeing local AI without NVIDIA
AMD
AMD Stumbles Again in Local AI Inference — Software Generation Behind Hardware
Lemonade drops AMD ROCm for Vulkan — benchmarks show ~40x speedup, exposing AMD's chronic gap: strong silicon, weak software versus Nvidia's CUDA.
Qwen3.5
Qwen 3.5 35B Benchmarks: Vulkan vs ROCm on AMD Strix Halo
Vulkan wins token generation (~57.5 t/s) while ROCm leads prompt processing (~1052 t/s) on AMD Ryzen AI MAX+ 395.
Apr 83 min readopcnew Editorialreddit.com/r/LocalLLaMA