Developer MAAN released OpenDLSS-NR, an open-source Vulkan reimplementation of NVIDIA's DLSS 5 Neural Rendering that runs on RTX 40 GPUs and includes a WebGPU implementation for web browsers. The project recreates the original neural network with bit-exact results and includes optimized kernels for Tensor Cores, though it requires users to supply NVIDIA's proprietary model weights separately.
This article explains how to efficiently multiply matrices when their parameters are distributed across multiple accelerators like TPUs and GPUs. It introduces a notation system using device meshes and sharding assignments to describe how tensor dimensions are partitioned across physical hardware, enabling scalable training of large language models.