Gpu Computing RSS Feeds

Directory1 gpu computing feeds · 32 articlesShow all languages →

Latest

Gpu Computing articles

Swordfish: Blackwell-Native GEMM Kernel Family for GPU Inference (PR #1707)

Swordfish is a Blackwell-native (sm100 and sm110) GEMM kernel family descended from the same lineage as Marlin and Machete. Supported modes: symmetric GPTQ int4 (uint4b8) and int8 (uint8b128), AWQ ...

Codetechnologymachine learningprogramming

github.com·Twitter AI (high signal)·1d ago·8 min read

Unsloth Dynamic NVFP4: A Guide to 4-Bit Inference on NVIDIA Blackwell GPUs

technologymachine learning

Shared by @UnslothAI ↗

unsloth.ai4d ago

Flash-MSA: Open-source sparse attention training kernels for million-token GPU training

Insighttechnologymachine learning

nanduruganesh.github.io5d ago

DFlash Speculative Decoding Boosts NVIDIA Blackwell Inference Performance Up to 15x

Insighttechnologyprogramming

Shared by @PyTorch ↗

developer.nvidia.com8d ago

Achieving 232x Speedup on GPU QR Factorization: A Contest Retrospective

Table of Contents Intro Contest in short Problem intro Why this problem is auto-research-able Learning Enough to Ask Better Questions (Optional) Math f...

Insighttechnologyprogrammingmachine learning infrastructure

sankalp.bearblog.dev·Twitter AI (high signal)·9d ago·23 min read

PyTorch Monarch Distributed Training Now Available on AMD GPUs via ROCm

Training state-of-the-art large language models (LLMs) with billions of parameters requires distributed training across hundreds or thousands of GPUs. At this scale, hardware failures are not exceptional events—they are expected. A single GPU memory error

Insighttechnologymachine learningprogramming

pytorch.org·Twitter AI·11d ago·9 min read

FP8 Tensor Cores Enable FP64-Equivalent 3-D FFT on NVIDIA Blackwell Ultra via Ozaki-Bailey Decomposition

NVIDIA's Blackwell Ultra (B300) cuts FP64 vector throughput to ~1.3 TFLOPS per GPU, roughly 30x below B200 and well below the level at which bandwidth-limited FP64 workloads stay memory-bound. The Ozaki Scheme II framework recovers FP64-equivalent through

Insighttechnologyprogramminghigh-performance computing

arxiv.org·Tech by Flipboard·15d ago·2 min read

Dynamic GPU Capacity Controller Reallocates Idle Production GPUs to Research During Off-Peak Hours

Insighttechnologyprogramming

Shared by @c_valenzuelab ↗

runwayml.com15d ago

ZLUDA Q1 & Q2 2026 Update: PhysX pre-alpha, Blender textures, and improved Windows support

Newstechnologyprogramming

vosen.github.io18d ago

TileMaxSim: IO-Aware GPU Kernels Achieve 80% HBM Bandwidth for Multi-Vector Retrieval Scoring

Multi-vector retrieval models such as ColBERT achieve state-of-the-art accuracy through fine-grained token-level MaxSim scoring, yet existing GPU implementations leave most hardware performance unused. We give a roofline analysis of MaxSim on modern GPUs

Insighttechnologymachine learning

Shared by @reachsumit.com ↗

arxiv.org22d ago

The Critical Role of GPU Kernel Quality in Machine Learning System Performance

Insighttechnologymachine learning

Shared by @Tim_Dettmers ↗

mlc.ai24d ago

wasi:webgpu - A WASI Proposal for GPU Access in WebAssembly

Codetechnologyprogramming

github.com1mo ago

Building high-performance expert-parallel dispatch and combine kernels for MoE LLM inference

Insighttechnologyprogramming

fergusfinn.com1mo ago

APEX4: Platform-Dependent W4A4 LLM Inference via Intra-SM Compute Rebalancing

W4A4 quantization promises full utilization of INT4 Tensor Cores, yet group dequantization overhead on CUDA Cores has driven existing systems to mixed-precision fallbacks. We present the first systematic study of how intra-SM compute balance governs this

Newstechnologymachine learningprogramming

arxiv.org·Mental Health Matters·1mo ago·2 min read

NVIDIA Announces DLSS 4.5 Support Coming to Blender 5.3

Newstechnologydesign

Shared by @80lvl-bot.bsky.social ↗

80.lv1mo ago

Rotary GPU: Enabling Large Mixture-of-Experts Models on Consumer Laptop GPUs with Limited Memory

Large language models have achieved remarkable capabilities through scaling, and this paper does not challenge that. It instead investigates a different question: once large models already exist, can they become more accessible to environments with substa

Insighttechnologymachine learning

arxiv.org1mo ago

Unsloth and NVIDIA Partner to Accelerate LLM Fine-Tuning by 20%

Newstechnologyprogramming

Unsloth - Train and Run Models Locally2mo ago

Systalyze's Utilyze Tool Reveals True GPU Compute Utilization in AI Workloads

Insighttechnologyprogramming

systalyze.com2mo ago

AMD1

Swordfish: Blackwell-Native GEMM Kernel Family for GPU Inference (PR #1707)

Unsloth Dynamic NVFP4: A Guide to 4-Bit Inference on NVIDIA Blackwell GPUs

Flash-MSA: Open-source sparse attention training kernels for million-token GPU training

DFlash Speculative Decoding Boosts NVIDIA Blackwell Inference Performance Up to 15x

Achieving 232x Speedup on GPU QR Factorization: A Contest Retrospective

PyTorch Monarch Distributed Training Now Available on AMD GPUs via ROCm

FP8 Tensor Cores Enable FP64-Equivalent 3-D FFT on NVIDIA Blackwell Ultra via Ozaki-Bailey Decomposition

Dynamic GPU Capacity Controller Reallocates Idle Production GPUs to Research During Off-Peak Hours

ZLUDA Q1 & Q2 2026 Update: PhysX pre-alpha, Blender textures, and improved Windows support

TileMaxSim: IO-Aware GPU Kernels Achieve 80% HBM Bandwidth for Multi-Vector Retrieval Scoring

The Critical Role of GPU Kernel Quality in Machine Learning System Performance

wasi:webgpu - A WASI Proposal for GPU Access in WebAssembly

Building high-performance expert-parallel dispatch and combine kernels for MoE LLM inference

APEX4: Platform-Dependent W4A4 LLM Inference via Intra-SM Compute Rebalancing

NVIDIA Announces DLSS 4.5 Support Coming to Blender 5.3

Rotary GPU: Enabling Large Mixture-of-Experts Models on Consumer Laptop GPUs with Limited Memory

Unsloth and NVIDIA Partner to Accelerate LLM Fine-Tuning by 20%

Systalyze's Utilyze Tool Reveals True GPU Compute Utilization in AI Workloads