A new rumor about AMD’s next-generation RDNA 3 GPU architecture has surfaced, with AMD software engineers releasing new patches for the new GFX11 architecture, also known as RDNA 3.
SEE GALLERY – 3 IMAGES
The new patch shows that AMD is working with its own instructions that can operate on arrays, with support for advanced artificial intelligence algorithms, to power its superresolution technologies on RDNA 3-based graphics cards, detected by @ 0x22h on Twitter. Indeed, this would be like Tensor cores on an NVIDIA GPU, hardware-based support for DLSS.
AMDGPU is a backend for AMD GPUs for the LLVM compiler library, which AMD employees update, and if you pay enough attention, you can find out what next-generation GPU architectures will bring. In this case, the Wave Matrix Multiply-Accumulate was added to the new GFX11 architecture, with GFX11 being the code name for AMD’s RDNA 3-based Radeon RX 7000 series GPUs.
The new Wave Matrix Multiply-Accumulate (WMMA) statement will work on arrays, without Keanu Reeves, rectangular arrays of tables that have numbers. The AI (artificial intelligence) and ML (machine learning) algorithms are based on WMMA for large sets of numbers, and it is not the first time that AMD has an architecture that supports array operations, as its architecture CDNA GPU supports it.
The AMD CDNA architecture has an instruction called MFMA (Matrix-Fused-Multiply-Add) which is different from WMMA (Wave Matrix Multiply-Accumulate) in terms of the format of supported arrays and output formats. Part of the code was published in AMDGPU, which points out that WMMA only supports 16x16x16 arrays, while it can come out in FP16 and BF16 data formats.
NVIDIA uses array multiplications for its deep learning operations through its Tensor Core architecture, with AMD’s new WMMA instructions working at a hardware level that enhances machine learning (ML) and DNN operations . Interesting, but totally expected. AMD has taken its FidelityFX Super Resolution (FSR) technology seriously, with FSR 1.0 evolving into FSR 2.0 just before the next-generation RDNA 3 GPU architecture is developed.
FSR 3.0 should be very interesting … and FSR 4.0 onwards … well.
NVIDIA is already well into DLSS 2.x and will unlock more magic than DLSS 3.0, while if you mix it up, they’re also about to launch their next-generation Ada Lovelace GPU architecture and GeForce RTX 40 series graphics cards. I’m sure you’ll be ready to rock and roll with Tensor Cores and updated DLSS 3.0 skills.