DeepSeek R1 1.5B running locally in the browser with WebGPU

This is cool, but something tells me the t/sec rate compared to native is still wildly off-base. How much less optimized is WebGPU compared to the current OpenCL/CUDA implementations?