Tag: #gpu-acceleration26 results
PyTorch Blog announces optimizations for diffusion models on NVIDIA Blackwell GPUs using MXFP8 and NVFP4 formats with Diffusers and TorchAO. These enable super-realistic image and video generation while addressing memory and compute constraints. This boosts adoption of diffusion models in AI workflows.
fastrad is a PyTorch-native radiomics library offering 25× speedup over PyRadiomics on GPU. It fully implements all 8 IBSI feature classes with device-agnostic tensor ops. Validated for accuracy against benchmarks and real datasets.
Moore Threads achieves systemic FP8 breakthrough, syncing with Nvidia globally as one of few domestic GPU makers. MTT S5000 delivers 1000 TFLOPS FP8 AI compute with 80GB VRAM and full precision support. MUSA platform enables compatibility with PyTorch, vLLM, and other AI frameworks.
Import AI 448 highlights ongoing AI R&D, including Bytedance's new agent capable of writing CUDA code for GPU tasks. It also covers advancements in on-device AI for satellites. The issue questions when the first major AI war might occur, paralleling Ukraine's drone war.
The cayleyR R package introduces a novel approach to solving permutation puzzles like TopSpin by detecting cycle intersections within Cayley graphs. It utilizes an iterative bidirectional search algorithm, enhanced by C++ performance optimizations and optional GPU acceleration.
NVIDIA CUDA 13.3 introduces native support for carryless multiplication, a critical primitive for cryptography and zero-knowledge proofs. This update bridges a long-standing performance gap between x86 CPUs and NVIDIA GPUs.
NVIDIA introduces GPU-accelerated Presto, designed to deliver peak performance for large-scale analytical query workloads. This integration leverages the GB200 NVL72 architecture to provide low-latency query results for users and AI agents.
Mozilla has merged initial support for Vulkan Video into Firefox, providing a new GPU hardware decoding path. This update aims to improve cross-platform performance and reduce reliance on VA-API.
AMD has announced that its latest AI-driven upscaling technology, FSR 4.1, will be extended to older GPU architectures. This update includes support for the RX 7000 and RX 6000 series, as well as Steam Machine hardware.
ibu-boost is a new gradient-boosted tree library applying a screening transform to absolutely reject poor splits, eliminating the need for min_gain_to_split tuning. It supports non-oblivious and oblivious trees, MSE/binary loss, missing values, and Triton GPU kernels for speedups. Benchmarks on California Housing show 12% RMSE gap to LightGBM but 3x GPU acceleration.