For several generations of gaming and specialized NVIDIA GPUs, special tensor blocks have been used to work with matrix data. These blocks allow you to accelerate deep learning calculations, which expands the capabilities of NVIDIA solutions in specialized tasks. These blocks have also found use in the gaming segment, since Deep Learning Super Sampling (DLSS) intelligent scaling technology uses calculations using tensor blocks. Now these blocks can get next-generation AMD GPUs. This is indicated by the data in the updated LLVM libraries of the AMDGPU driver.
For the architecture codenamed GFX11 (RDNA3), Wave Matrix Multiply-Accumulate (WMMA) technology has been added with support for special instructions for matrix calculations. WMMA supports 16x16x16 matrices and can output data in FP16 and BF16 formats. This is not the first architecture to support matrix operations. The CDNA architecture for Instinct MI200 accelerators already supports Matrix-Fused-Multiply-Add (MFMA) instructions. But Instinct accelerators are a data center product, and in the case of RDNA3, we are talking about mass solutions. With such a hardware upgrade, FidelityFX Super Resolution technology can be developed, which will become a direct analogue of NVIDIA DLSS.
Source:
videocardz.com
