Dev Tools · 1h ago
TensorSharp adds Vulkan backend, matches llama.cpp on GPU inference
TensorSharp, a .NET tensor library, now supports Vulkan GPU acceleration via an external GGML backend. Benchmarks on Gemma 4 models show TensorSharp's decode throughput is within 5-10% of llama.cpp on Vulkan, and prefill throughput is competitive. The developer is seeking AMD GPU testers for broader compatibility.
Meridian48 take
The performance parity with llama.cpp is notable, but the reliance on an external backend and limited GPU testing means this is early-stage for production use.
tensor-libraryvulkan-backend