Tag: #vulkan6 results
Benchmarks compare Nvidia RTX 3090 and Intel Arc Pro B70 using llama.cpp on Vulkan and SYCL backends across multiple GGUF models. Arc Pro B70 shows 71% slower prompt processing and 53% slower token generation on average versus RTX 3090. SYCL backend improves some Arc results over Vulkan.
ZINC is a new LLM inference engine written in Zig, enabling 35B models on $550 AMD GPUs via Vulkan. It loads GGUF models, achieves 7.1 tok/s on RDNA4, and addresses AMD consumer GPU gaps. Repo at github.com/zolotukhin/zinc.
A new pull request aims to make Tensor Parallelism viable when using the Vulkan backend in llama.cpp. This update is a significant step toward improving multi-GPU performance for non-Nvidia hardware.
User benchmarks and consolidates homelab from three models to a single 122B Qwen3.5 MoE on Strix Halo, achieving 27.4 tok/s and top scores. IQ3 quantization matches Q4_K_M performance at half VRAM. Handles multiple apps like email, finance, and cameras concurrently.
Mesa 26.3 has merged AMD’s GFX1171 target, adding support across the RADV Vulkan and RadeonSI OpenGL drivers. The change prepares Linux’s open-source graphics stack for an undisclosed RDNA 4m integrated GPU variant.
User reports excellent results using Q5_K_XL quantized Qwen3.6-35B-A3B on AMD RX 7900 with llama.cpp Vulkan server. Integrated with VS Code Copilot for generating full React apps and Playwright tests from large prompts. Shares detailed server command, sampling params, and chat config.