A new optimization for FlashAttention-4 (FA4) decoding on NVIDIA Blackwell GPUs overlaps softmax and matrix multiplication operations by using spare tensor memory (TMEM) in a ping-pong scheme, achievi…
A new technical blog post by Iaroslav Elistratov presents a visual guide to building a B200 attention kernel from scratch in CUDA and PTX, achieving 94.4% of FlashAttention-4 performance on 4K, 8K, an…
Nvidia's AVO agent, powered by Anthropic's Claude Opus 5, achieved a 100% success rate on the ARC-AGI-3 benchmark, completing all 183 levels across 25 environments on August 21 while using 12% fewer a…
NVIDIA's Agentic Variation Operators (AVO) architecture achieved a 100% score on the ARC-AGI-3 benchmark, demonstrating a frontier-level general-purpose architecture for long-horizon autonomous agents…
Thinking Machines Lab released Inkling, a new multimodal mixture-of-experts model for token-efficient reasoning and native multimodal understanding, and Together AI made it available on its inference …
AI agent infrastructure is maturing but facing reliability challenges, as FlashAttention-4 achieves 71% utilization on NVIDIA's Blackwell B200 GPUs while agentic AI systems grapple with architecture q…
An analysis of frontier AI model release cadence shows Anthropic and OpenAI accelerating their release rates since 2023, while Google, Meta, and DeepSeek have not. The pattern is consistent with a for…
Researchers from Princeton University, Together AI, Meta, Colfax Research, NVIDIA, and Georgia Tech introduced FlashAttention-4, an algorithm and kernel co-design that optimizes attention for Blackwel…
Modal AI engineers Charles Frye and David Wang optimized FlashAttention-4 for large language model inference, focusing on decode-heavy workloads dominated by memory bandwidth-limited token generation.…