Annapurna Labs

The biggest barrier to adopting new AI hardware is the months spent rewriting and revalidating code that already works. Announcing TorchNeuron, which lets PyTorch code run natively on AWS Trainium chips with a single configuration change. No rewrite, and the tools developers already use to test, debug, and scale their models all work on Trainium without modification. Startups and enterprises in robotics, video, world models, self-driving cars, speech, drug discovery, LLMs, and finance are already building on it. TorchNeuron will be open sourced at the PyTorch Conference on October 20: https://amzn.to/4icd1Ut

The chips powering the biggest names in AI? Built at Amazon. ⚡ Anthropic runs Claude on 1M+ Trainium2 chips. OpenAI committed 2 gigawatts of Trainium capacity. Meta committed tens of millions of Graviton cores to agentic AI. 98% of our top 1,000 EC2 customers run on Graviton. 💡 It started with a bet a decade ago — design chips for cloud workloads instead of relying on general-purpose hardware. Today, that bet is paying off with triple-digit YoY growth. 🔧 The engineers behind this work at the intersection of hardware and software — designing silicon from the ground up, then building systems at unprecedented scale. 🧠 Trainium3 UltraServers pack 144 chips into a single system. Trainium4 is already in development. "One of the ways that we can give ourselves an advantage is by using our deep investments in chips to deliver both performance and cost efficiency that will allow us to differentiate our model development." — Peter DeSantis, SVP, Amazon AI Models, Chips, & Quantum Computing 📖 Read the full story: https://lnkd.in/gRbbCjnb

In 2015, Amazon acquired a small chip design company called Annapurna Labs. The AI boom was still years away. The bet: purpose-built silicon for cloud workloads would deliver better performance at lower cost. 🎯 Fast forward to today — Amazon's custom chip business has reached a $25 billion annual revenue run rate, growing triple-digit percentages year over year. 📈 Graviton chips now power more than 130,000 customers, including 98% of the top 1,000 Amazon Web Services (AWS) EC2 customers. I've spent the past couple of years taking reporters inside our Austin chip lab, watching engineers weld microscopic circuits and power on silicon for the first time. This new About Amazon piece tells the full arc — from that 2015 acquisition through Graviton, Nitro, and Trainium to a business Jeff Bezos now calls Amazon's next "pillar." ⚡ Proud to help tell this story every day. 🙏 Read the full piece: https://lnkd.in/g-n-g5D7 #Amazon #AWS #Silicon #cloud #TechInnovation #AI #Chips Isaac Schultz Nafea Bshara Ron Diamant Kristopher King Mark Carroll Colin Brace

Graviton5 is now generally available. Amazon's fifth-generation custom silicon introduces a four-chiplet architecture with 192 cores, custom die-to-die connectivity delivering up to 420 GB/s bandwidth, DDR5-8800 memory support, and PCIe gen6 interconnects — delivering 25% better performance than Graviton4 for general-purpose and agentic AI workloads.     It also debuts the Nitro Isolation Engine, the first formally verified cloud hypervisor, which enforces VM isolation with mathematical precision: https://amzn.to/4eAlnUY

AI researchers, if you are heading to NeurIPS this week you won’t want to miss the hands-on workshop about playing chess on AWS Trainium! Emily Webber will be hosting live tournament for you to explore trade-offs in reasoning, fine-tuning, and realtime gameplay. We’ll also discuss some of the latest AI research from Amazon. Join us at 12 pm on Tuesday December 2, Upper Level Ballroom 20AB. https://lnkd.in/gEmhq5jV https://lnkd.in/gBMTVxax #AWSTrainium

re:Invent is coming this week! I will be running two AWS Trainium workshops: - [AIM309] Building LLMs with AWS Trainium - [AIM318] Fine-tune Hugging Face LLMs on AWS Trainium with Optimum Neuron Please join one of the workshops to learn how to work with AWS Trainium2-powered instance for LLM fine-tuning and deployment. https://lnkd.in/gd_MMs2m https://lnkd.in/g6YzycmQ We are also showcasing some very exciting demos. Visit the AWS Trainium Booth at the AWS Village in the Expo Hall. https://lnkd.in/gBdvP8Hg 日本語でも対応しますのでお声がけ下さい。 #aws #reInvent

Chess, anyone? This year at NeurIPS I am thrilled to announce a hands-on workshop about playing chess on AWS Trainium! I'll be onsite in the 2000-person ballroom on Tuesday Dec 2, with an amazing team. Chess is a great learning environment for many aspects in machine learning. The game itself is well handled by CPU-based gaming engines, but the ability of LLMs to understand these discrete rules and common strategies is vastly under-explored. In addition, there's a natural pair with both reasoning and chat. Imagine training your own small language model to not just play chess, but also explain it's moves to you and have a dialogue with you about this. This gives you the ideal environment to grow your skills, both as a data scientist and as a chess enthusiast. Not to mention have a blast 🎸 We're bringing all of this to you and more at NeurIPS, and I couldn't be happier to team up with Rocky Yu at AGI House, Sharada Prasanna Mohanty at AI Crowd, Henry Yin, and the stellar builders at AWS with Jim Burtoft Niithiyn Vijeaswaran Megan McKenzie Josh Longenecker Ruchi Bhatia Armin A. and so many more. Space is limited - but if you're already attending NeurIPS you can jump into registration now while it's still open. https://lnkd.in/exVc8Qim Attending re:Invent and want to try it out? Don't worry, we'll have some options for you in December. Stay tuned ♟️.