We are releasing a major NATTEN upgrade that brings you new Hopper & Blackwell sparse attention kernels, both capable of realizing Theoretical Max Speedup: 90% sparsity -> 10X speedup. Thanks to the great efforts by Ali Hassani & @NVIDIA cutlass team! natten.org

Exclusive: Ex-Meta AI leaders debut an agent that scours the web for you in a push to ultimately give users their own digital ‘chief of staff’ Devi Parikh Abhishek Das @DhruvBatraDB fortune.com/2025/06/10/exc…

We're excited to launch Scouts — always-on AI agents that monitor the web for anything you care about.

It's here! Introducing Scouts by Yutori. Scouts is like having a team of agents monitoring the web for information that matters to you. We're letting more users in everyday. Join the waitlist!

Scouts by Yutori. AI agents that monitor the web for things you care about. So you can focus on the meaningful things in life and experience a bit more yutori.

Scouts from Yutori keep tabs on things you care about so you never miss an important update! An update like AMD securing a big win in the AI chip market :)

Oh wow, the newly released nccl finally started to use fp32 accumulation for reduction ops with half precision inputs! This is so important! Thank you NCCL team! github.com/NVIDIA/nccl/co… I'd imagine we will see this version in pytorch>=2.8 I hope, unless you build your own nccl.

the "compute bound" workload in question

hitting ^C^C^C^C^C^C^C^C to get the shell back after the 73rd torchrun run in the debugging session crashes

Cosmos-Predict2 meets NATTEN. We just released variants of Cosmos-Predict2 where we replace most self attentions with neighborhood attention, bringing up to 2.6X end-to-end speedup, with minimal effect on quality! github.com/nvidia-cosmos/… (1/5)

Sparse Attention is now pushing World Foundation Models to the Speed of Light! Attention powers modern AI (Transformers, ViTs, DiTs), and Sparse Attention is the next frontier. Neighborhood Attention (NA) is the first multidimensional sparse attention infrastructure that: -

Watch my talk about NATTEN on GPU MODE this Saturday at 3PM ET / noon PT. I'll go over all the exciting new features we shipped very recently, especially our Hopper and Blackwell FNA kernels, now speeding up video / world models by up to 2.6X e2e! youtube.com/watch?v=mF_H_J

Check out Ali's talk tomorrow on GPU MODE if you breathe GPUs! This is our 2nd GPU MODE talk — last time we unveiled Distributed GEMM: a CUTLASS-based Tensor Parallelism implementation that helps push NVL-based AI systems to the next level - transforming a network of GPUs into

Not sure why but we have lots of new followers! FFmpeg makes extensive use of hand-written assembly code for huge (10-50x) speed increases so we are providing assembly lessons to teach a new generation of assembly language programmers. Learn more here: github.com/FFmpeg/asm-les…

i truly hate to say it but i have seen normalization kernels in SF that achieved higher MFU than matmuls in LA

5, 5-high, 5-high-fast, 5-low, 5-low-fast

God bless Ali Hassani for shipping reference kernels with NATTEN.

Legend has it gym memberships are particularly hard to cancel here because nobody knows how to exit

how it feels to know the location of a native CUDA kernel in pytorch