←
Lambda2,077 posts
Lambda avatar

Lambda

@LambdaAPI

The Superintelligence Cloud

◷ Joined July 2012▧ 567 Photos and videos
Lambda avatar
Lambda@LambdaAPI·Oct 29

Others announce. We deliver. Proud to be recognized by enterprises, hyperscalers & frontier labs.

0:45
Lambda avatar
Lambda@LambdaAPI·3h

Cooler optics, steadier signals, lower latency: here’s what we measured on our CPO NVIDIA Photonics switches. Thanks, @NVIDIA. #SC26

AI factories need networks that keep thousands of GPUs working together efficiently and reliably. @LambdaAPI is deploying NVIDIA Photonics co-packaged optics switches to support AI infrastructure at scale with 10% lower latency, 15x higher reliability, and lower switch temperature compared to conventional pluggable-based switching. #SC26

0:43
NVIDIA AI Infrastructure avatar
⟳ Lambda retweeted
NVIDIA AI Infrastructure@NVIDIAAIInfra·Nov 17

AI factories need networks that keep thousands of GPUs working together efficiently and reliably. @LambdaAPI is deploying NVIDIA Photonics co-packaged optics switches to support AI infrastructure at scale with 10% lower latency, 15x higher reliability, and lower switch temperature compared to conventional pluggable-based switching. #SC26

0:43
Lambda avatar
Lambda@LambdaAPI·Nov 16

In Ian Buck's special address at #SC26: Lambda completes the world’s first large-scale deployment of co-packaged optics @NVIDIA Photonics switches.

Post attachment
Lambda avatar
Lambda@LambdaAPI·Nov 16

The Flux.1 speedup breaks down cleanly. 1.27× came from a newer NVIDIA software stack (NVIDIA CUDA™ 13.4.1, PyTorch NVIDIA 26.08). Another 1.18× came from an optimized configuration (lr 2.5 × 10⁻⁴, global batch 1024, 1600 warmup steps). Our unofficial Flux.1 result was 1.12× faster than AMD's official 64 MI355X submission on the same workload. Code and configuration are public: github.com/LambdaLabsML/lamb…

Post attachment
Lambda avatar
Lambda@LambdaAPI·Nov 16

GPT-OSS-20B is one of MLPerf's first MoE training benchmarks. MoE models activate only a subset of parameters per token, and MoE is one of the most widely used architectures behind LLM applications today, so this is a useful reference if you are building reasoning or agentic applications on MoE architectures. Flux.1 is Lambda's first multimodal workload through the MLPerf pipeline. Text-to-image models stress a system differently than language models. They're increasingly used to generate synthetic data for robotics and autonomous vehicles.

Post attachment
Lambda avatar
Lambda@LambdaAPI·Nov 16

MLPerf Training v6.1 results are out! On 64 @NVIDIAAI Blackwell Ultra GPUs in a single NVIDIA GB300 NVL72 system, GPT-OSS-20B trained in 19.697 minutes. That is 2.86× faster than 64 NVIDIA H100 GPUs last round (Cisco) and 1.37× faster than 64 NVIDIA Blackwell GPUs (CoreWeave). We also ran our first multimodal benchmark. In an unofficial reproduction, Flux.1 trained in 26.36 minutes on the same system, 1.50× faster than the last round's fastest on that hardware. New hardware shortens training, and better software keeps finding speed on hardware you already have.

Post attachment
Lambda avatar
Lambda@LambdaAPI·Nov 16

At an AI campus scale, the network puts the “super” in supercomputer. Lambda has completed the world’s first large-scale deployment of co-packaged optics (CPO) networking, connecting 10,000+ NVIDIA Blackwell Ultra GPUs using NVIDIA Photonics networking switches. Compared with conventional pluggable optics, we observed up to 10% lower latency, up to 5x better power efficiency, and error rates that held steady as the fabric grew. That means more of the power budget goes to compute, and fewer interruptions for teams running long training jobs. Read the announcement: lambda.com/blog/worlds-first… @NVIDIA @NVIDIAAIInfra #SC26

Post attachment
Lambda avatar
Lambda@LambdaAPI·Nov 13

Lambda is on the inaugural AI 100 from @NewcomerMedia and @BessemerVP. Congrats to all the honorees. Many of them are Lambda customers, and we’re glad to be the compute behind their work. Full list: ai100.newcomer.co

The AI 100 is live 🏆 Our inaugural list of the top 100 private AI companies just dropped on @NewcomerMedia. Honorees were evaluated on market leadership, estimated valuation, operating metrics, and people & culture, ranging from foundation models and AI infra to robotics and beyond. Congratulations to the companies making AI history! See it here: ai100.newcomer.co

Post attachment
Lambda avatar
Lambda@LambdaAPI·Nov 13

TurboPairFormer: open-source GPU kernels that deliver up to 1.73× faster OpenFold3 GPU computation per optimizer step than the original Triton backend, with more accurate gradients and reproducible results in fixed-input tests. Triangle attention scales cubically, so doubling the training crop makes it ~8× more expensive. Existing backends each trade away speed, precision, or reproducibility. Researchers from Lambda, Stevens Institute of Technology, and the OpenFold team built TurboPairFormer to address all three on NVIDIA Hopper GPUs: • Up to 1.73× faster than OpenFold3 Triton • Lowest mean Q/K and pair-bias gradient errors vs an FP64 reference • Bitwise-identical gradients across all four types (Q, K, value, pair-bias) in all 600 fixed-input cases Supports H100 and H200. On H200, 1.54× faster than OpenFold3 Triton on 768 pixel images (14.67 vs 22.53 sec/step). Read the blog for the full method and results lambda.com/blog/faster-more-…

Post attachment
Lambda avatar
Lambda@LambdaAPI·Nov 12

For most institutions, sovereign AI has meant a hard choice: pre-train your own model, a billion-dollar effort, or fine-tune one somebody else controls. Thomson Reuters and researchers at Imperial College London took a different route. They post-trained an open model on TR's own data and got most of the benefits of pre-training from scratch. Halfway through, the team moved its compute to Lambda and saw nearly 10x faster evaluation iterations. "If you outsource defining what good looks like to somebody else, you've then outsourced your AI strategy." — Nabeel Seedat, Thomson Reuters Read the story: lambda.com/customer-stories/…

Post attachment
Lambda avatar
Lambda@LambdaAPI·Nov 11

A nightly cleanup job doesn't need full admin access. In Lambda Cloud IAM, give it a service identity with InstanceReader and InstanceTerminate scoped to one workspace. Rotate its API key while the job keeps running. Workspaces hold resources. IAM controls what identities can do with them. The demo covers scoping permissions per workspace, giving automated jobs their own service identities and API keys, and managing shared permissions through groups. API key rotation keeps two keys active at once. Verify the replacement before retiring the old one. The job keeps running without a blanket admin role. View Effective Access brings together direct and group-inherited roles across workspaces, so you can audit what an identity can actually do before it runs. Watch the demo and see it all yourself:

3:33
Lambda avatar
Lambda@LambdaAPI·Nov 10

Burkov's new book takes you from the Policy Gradient Theorem to working PyTorch code. You implement REINFORCE from scratch, then move to PPO and Deep Q-Networks, all on the same rocket-landing task so you can compare how each algorithm learns and behaves. It also covers RLHF and verifiable rewards for language models. Lambda is sponsoring students with cloud credits to run the code.

BURKOV@burkovNov 10

TL;DR My new book is now available for purchase. As you probably know, in February 2025, I resigned from my full-time job to become a full-time author. At the same time, I started working on a new book in The Hundred-Page Books series. Then vibe coding happened, and I couldn't resist trying to build something this way. This resulted in ChapterPal, my AI-assisted reading and learning platform, which I've built without writing a single line of code. Today, with tools like Codex or Claude Code, it doesn't seem surprising, but when I started almost two years ago, I copied and pasted code back and forth to the chatbot window. It hallucinated solutions, rewrote what was working, refused to print the entire code, etc. It was a hell of an experience. Last spring, ChapterPal's maintenance was automated enough to allow me to return to writing, which I gladly did. And today, 20 months after I started, I finally finished my new book: The Hundred-Page Deep Reinforcement Learning Book. Before sending the manuscript to print, I took a two-month break so that my brain could read it as if for the first time. For the entire last week, I was reading my book and I got chills so many times that I'm sure it's a very good book. If you have been waiting for it, thanks for being patient. I hope you will see that the waiting was worth it. Today, the book is available on Amazon as a paperback, hardcover, and Kindle edition, and on Leanpub as a PDF. I'm currently preparing another hardcover edition, with different dimensions, for the rest of the online stores. For those who have my previous hundred-page books in their libraries, if you want the new hardcover book to be the same size as the previous two, you should probably wait for this non-Amazon edition to be published. I will announce this separately. A couple of words about the book. I liked the structure of my previous hundred-page book (the one on language models), where each chapter introduces a new algorithm that improves the previously introduced algorithms in some way. This makes the textbook read like a story rather than a collection of facts. The Hundred-Page Deep Reinforcement Learning Book builds the subject step by step, beginning with reinforcement learning fundamentals and the Policy Gradient Theorem before developing REINFORCE, actor-critic methods, Generalized Advantage Estimation, and Proximal Policy Optimization (PPO). It then introduces off-policy learning and derives Deep Q-learning (DQN) from first principles. The final chapter shows how reinforcement learning applies to language models, including learning from human feedback and verifiable rewards with PPO and GRPO (remember DeepSeek R1). The book comes with complete PyTorch implementations (on GitHub), the mathematical foundations needed to understand its content (in the appendices), and pointers for further study (in the last chapter). Additional learning resources are available on the book's wiki at thedrlbook.com/wiki. The book was read and endorsed by several people I deeply admire. The back cover contains endorsements from Long Ouyang, a co-creator of InstructGPT and ChatGPT at OpenAI and Stephen Balaban, a co-founder and CTO at Lambda. You can find all the endorsements at the book's website: thedrlbook.com. I hope you enjoy this new release in the series as much as you enjoyed the other two. I can't wait for the readers' reviews to start coming. (The Dark Edition is coming soon!)

Post attachment
Lambda avatar
Lambda@LambdaAPI·Nov 10

GPU-accelerated simulation is the bridge: embodied AI benchmarks in sim, then sim-to-real transfer for real robot policy learning. Three challenges, BEHAVIOR, RoboEval, RoboChess, make that pipeline concrete. Nov 12, 8:30 AM CST, Room 212, JW Marriott, Austin.

🤖 Excited to co-organize the CoRL 2026 Workshop: From GPU-Accelerated Simulation to Scalable and Generalizable Real Robot Policy Learning! Join us to explore how GPU-accelerated simulation, embodied AI benchmarks, and sim-to-real transfer can advance scalable robot learning. 🏆 Featuring three challenges: BEHAVIOR, RoboEval, and RoboChess, alongside invited talks from researchers at NVIDIA, CMU, Harvard, and UCLA. 📅 November 12, 2026 | 8:30 AM – 12:30 PM CST 📍 Room 212, JW Marriott, Austin, TX 🔗 corl2026-realworld-robot-lea… #CoRL2026 #PhysicalAI #Robotics #EmbodiedAI @LambdaAPI @nvidia @Stanford @RuohanZhang76 @jiajunwu_cs

Post attachment