Activity
6K followers
-
Sridhar Lakshmanamurthy shared thisRegister now🚀! This is an incredible opportunity to learn about how to shape the future of #CloudTPU. Vijay Degalahal and Naren Vallepalli have assembled an extraordinary team, you don’t want to miss this event ! 🎉Sridhar Lakshmanamurthy shared thisHello Everyone, On October 1, I’ll be joining Naren Vallepalli to deliver the opening keynote for our Spotlight on Google Cloud Silicon virtual event. We’ll be sharing our vision for the future of AI accelerators and how we co-design across hardware, software, and ML to make it happen. I want to make sure our keynote addresses what is top of mind for you. What are you most curious about regarding AI architecture, TPUs, or the challenges of hyperscale computing? Drop your questions in the comments below, and we will try to cover them! If you are a passionate silicon engineer who is excited about bringing your ideas to impact, come join the conversation! RSVP here: https://goo.gle/4hhF44b PS: We have limited spaces, so please secure your registration by September 30, 2026. Karthik Ramaswamy, Manjunatha Prabhushankar, Sunkari Sasidhar, Shantanu Samant, Hariprasad Gangadharan Puneet Shah #GoogleCloud #CloudSilicon #HardwareEngineering #LifeAtGoogle #GoogleIndia
-
Sridhar Lakshmanamurthy shared thisIt was truly an honor to have the opportunity to share the stage again with the legend, Norm Jouppi, at Hot Chips 2026. We discussed the trade-offs involved in building two TPU systems this year: 8i for inference and 8t for training. A huge shout-out and thank you to the entire TPU team for your creativity, tenacity and dedication. 👏 . You rock! 🎉 🚀
-
Sridhar Lakshmanamurthy shared thisIncredible opportunity to join the TPU family and build the future of ML Compute for the entire industry! #CloudTPU 🚀. And Vijay Degalahal is an awesome leader! Come join us!Sridhar Lakshmanamurthy shared thisHow do we scale AI compute to meet the demands of tomorrow's ML models? We build faster, smarter, efficient accelerators! Make the best AI Accelerator Better!! 🚀 Our team at Google Cloud India is looking for TPU Architects !! If you live and breathe Si Architecture have expertise in: ✅ Scale-Up & ✅ Scale-Out ✅ ML Hardware-Software Co-design ✅On die fabrics for massive BW scaling! ✅ Developing the most dense and power efficient compute ...then I want to hear from you! Let's build the future of AI infrastructure together. Join Google Cloud TPU Architecture in India. Multiple openings! Pls apply on the portal! Ravi Iyer Bharat Daga Ravi Krishnan Venkatesan Mahesh Natarajan Pradeep Kumar Janedula Shrikul Joshi Naren Vallepalli Hariprasad Gangadharan Praveen KS Varsha R. Surya Chongala Anand Kumar RBharat Daga Aditya Yanamandra
-
Sridhar Lakshmanamurthy shared thisPlease register for this event if you are curious to know more about TPU chip and system codesign! 🚀Sridhar Lakshmanamurthy shared thisJoin us on June 23rd for our virtual event: "Vertical Integration & Custom Silicon: Engineering Google's AI Future." Hear from Sridhar Lakshmanamurthy, Senior Director of TPU Chip Architecture, followed by the opportunity to join a breakout session lead by Google leaders across the organization. Areas include RTL/DV, Physical Design, Infrastructure Tools & Methodology (yours truly!), and Post-Silicon Validation. Date: Tuesday, June 23rd Time: 12:00 PM – 1:00 PM PST Format: Virtual Whether you are actively looking or just curious about our vertically integrated approach—from TPU architectures to full-system co-design—we hope you’ll join us. Register now to secure your spot: https://goo.gle/3Q9mnGz
-
Sridhar Lakshmanamurthy shared thisBeing a coauthor on this paper alongside computer architecture legends Dave Patterson and Norm Jouppi and getting to work on TPU architecture with them and other incredibly talented yet kind human beings is definitely a highlight of my career! 🙏 Go #CloudTPU #Ironwood #TPU8i #TPU8tSridhar Lakshmanamurthy shared thisWe just put a paper on arXiv that will appear in "IEEE Micro" that reviews 5 generations of TPUs up through Ironwood, showing the remarkable longevity and scalability from the initial training TPU a decade earlier: https://lnkd.in/gmpBc95X The paper also shows how dramatically and quickly production deep neural network models change:MLPs were 60% a decade ago & now 11%; RNNs were 29% & now ~0%; Transformers were 0% then & 74% now. Amazingly , the TPU architecture gracefully handled the transitions, just changing speeds and number of components each generation but keeping the same underlying design.
-
Sridhar Lakshmanamurthy shared thisTimmy the TPU is here! Makes for a great swag! 😆 We are still hiring, come join us on this incredible adventure contact Rachael Tiss for details 🚀🚀Sridhar Lakshmanamurthy shared thisIn my career, I've had the privilege of working on a lot of incredible chips—but this is officially the first one that gets its own cartoon! #GoogleIO #TPU https://lnkd.in/d9REkm9j
-
Sridhar Lakshmanamurthy shared thisOf course we are serious. For the record, we now 2x more serious! 😆 Uri Frank thank you for your leadership! We are still hiring, come join us! #CloudTPU #TPU8Sridhar Lakshmanamurthy shared thisIn the past five years talking with dozens of candidates, the most frequent question I got was “ How serious is Google about developing their own silicon” Well - actions speak louder than words, I do not get that question anymore. Big announcements coming out of Cloud next: ✅ TPU 8t: The Training Powerhouse, optimized to shrink frontier model development from months to weeks. ✅ TPU 8i: The Reasoning Engine, built for ultra-low latency to power real-time agentic experiences. ✅ Custom Axion Arm-based CPU hosts, allowing us to optimize the entire system for maximum power efficiency. Want to join? We are Serious about silicon! https://lnkd.in/dZ-zRRQt #GoogleCloudNext #AI #TPU8 #Axion
-
Sridhar Lakshmanamurthy shared thisThank you Amin Vahdat and Jeff Dean for your leadership and support in codesigning these awesome TPUs! Your astute insights helped shape so many of the groundbreaking features for both TPU 8t and 8i.Sridhar Lakshmanamurthy shared thisAs a student of history, I always welcome the opportunity to revisit the incredible progress we have made toward architecting the future of computing for the agentic era. Jeff Dean and I have spent a lot of time together reflecting on both the path behind us but also the path ahead. So it was a real pleasure for the two of us to sit down with Ben Gilbert and David Rosenthal, hosts of the Acquired podcast last week at Cloud Next for a deep dive into the inflection points that defined Google’s AI infrastructure. We hit many of the highlights including the origin story of TPUs, how the capabilities grew generation over generation, and some of what we see for the future of AI infrastructure. You can watch the full video here: https://lnkd.in/dtJr4ewF
-
Sridhar Lakshmanamurthy shared thisReally exciting times for all the chip and system teams at #Google!! Next generation TPUs paired with Axion Arm- based host CPUs! All in-house silicon! 🚀 #CloudTPUSridhar Lakshmanamurthy shared thisWe are entering the age of agents, where models don’t just respond—they reason, execute workflows, and learn. To meet these new demands, I’m excited to announce the eighth generation of Google’s custom TPUs at Google Cloud Next, featuring two specialized architectures: TPU 8t: The Training Powerhouse, optimized to shrink frontier model development from months to weeks. - Scales to 9,600 chips and 2 PB of shared memory per superpod. - Delivers 121 ExaFlops of compute with 2.7x better price-performance than our previous generation, Ironwood. - Engineered for 97% "goodput" via automated fault detection and rerouting. TPU 8i: The Reasoning Engine, built for ultra-low latency to power real-time agentic experiences. - Breaks the "memory wall" with 3x more on-chip SRAM (384 MB) to keep active models entirely on-chip. - Features the new Boardfly topology, doubling interconnect bandwidth to 19.2 TB/s. - Provides 80% better performance-per-dollar for serving compared to Ironwood. For the first time, both chips run on our custom Axion Arm-based CPU hosts, allowing us to optimize the entire system for maximum power efficiency. We remain open-by-design, supporting PyTorch, JAX, and vLLM, while offering bare metal access for the first time. I’m eager to see how these systems empower our customers to build capabilities at the cutting edge of what's possible. Read more about TPU 8t and 8i in my blog post here. https://lnkd.in/g39Zm8BA #GoogleCloudNext #AI #TPU8 #CloudInfrastructure #MachineLearning
Sridhar Lakshmanamurthy commented on a post
1w
Congratulations Raj! 👏 Raj Yavatkar + Ram Sunder, now that is a DreamTeam!!
Sridhar Lakshmanamurthy commented on a post
4w
Congratulations 🎉 Sabareesh R. Uriel (Ben) Schafer Daniel Rosenband and the entire Waymo Compute team ! Truly impressive accomplishment! 👏
-
Sridhar Lakshmanamurthy reacted on thisHello world. We took a little break but we are back now.😊 I have been using it for awhile and it's good. Look forward to all of your takes. #Gemini4 #AISridhar Lakshmanamurthy reacted on thisToday we’re introducing Gemini 4 Argon, our next era of frontier intelligence. It delivers frontier performance in complex workflows across real-world software engineering, knowledge work, and cybersecurity defense with an industry-leading 1M token output limit. Gemini 4 Argon is rolling out to an initial cohort of cyber defenders through our Fairwind Program so they can leverage its full frontier-level cybersecurity defense capabilities. We'll continue to gather feedback from early testers as we iterate on guardrails before making Argon available to developers, enterprises, and consumers as soon as possible. Learn more → goo.gle/4AZLPRt
-
Sridhar Lakshmanamurthy liked thisSridhar Lakshmanamurthy liked thisFeatured Panel at the HYSTA 27th Annual Conference: "From Compute to Scale: The Future of AI Infrastructure" Moderated by Bill Jia, VP of Engineering, Core ML/AI at Google and HYSTA Board Member, the panel brings together: Ruoming Pang, Member of Technical Staff at OpenAI Banghua Zhu, Co-founder and CTO of RadixArk Ramine Roane, Corporate Vice President of AI Product Management & Ecosystem Development at AMD Junchen Jiang, CEO of Tensormesh, Co-creator of LMCache and Associate Professor at UChicago Register now to hear how these leaders see AI infrastructure evolving from compute to scale. 🎟 https://lnkd.in/gDni-ZbF 📍 Santa Clara Convention Center, Silicon Valley 🗓 Saturday, September 26, 2026 | 09:00 – 21:00 #HYSTA2026 #AIInfrastructure #Compute #MLSystems #SiliconValley #AIatScale
-
Sridhar Lakshmanamurthy liked thisSridhar Lakshmanamurthy liked thisNearly a year ago, I shared early news on Project Suncatcher, Google's moonshot to put AI in space. Why? Because the sun outputs ~100 trillion times as much power as humanity’s total electricity production, and we think sometime in the future, this will be something that humanity will take advantage of. So, we’ve begun the long ambitious journey to tap into that potential power by exploring the possibility of hosting scalable machine learning infrastructure in space. Today the New York Times wrote an in-depth story on Project Suncatcher -worth a read! https://lnkd.in/ecWjmr-V As with our other moonshots, like quantum computing or autonomous vehicles – there are many scientific and engineering challenges to solve for, and while we have many ideas, there is so much still to learn, research and innovate. So we are working towards that future vision via a series of milestones. And its worth adding that, as some other moonshots have shown, along the way we hope to learn and innovate in ways that may be useful in other contexts and applications including here on earth. So far we’ve been making steady progress with Project Suncatcher, including achieving some engineering and terrestrial milestones, and we’ll soon embark on our first test in orbit – sending Google’s TPUs into space to see how they perform. We will be aboard Transporter-18 rideshare mission with SpaceX. The question we’re asking early is deceptively simple: Can this hardware operate in space? With the initial mission, developed in partnership with Planet, we’ll be gathering in-orbit data on how TPUs can handle the physical stress of launch and the radiation and thermal extremes of space. In dawn-dusk sun synchronous low-Earth orbit, satellites can access near-constant sunlight, generating up to eight times more solar power than on Earth. Eventually, it may be possible to link together multiple constellations of satellites, allowing them to manage larger AI workloads while in orbit. Below, you can hear more from my colleagues on the Project Suncatcher team, who break down some of the science and engineering challenges we’re working through. And you can learn more about Project Suncatcher in this blog post: https://lnkd.in/e4GQdqdG
-
Sridhar Lakshmanamurthy liked thisSridhar Lakshmanamurthy liked thisIt was a pleasure sharing Google’s vision and approach for rapid AI growth at today’s Data Center World Power conference. Our progress is built on rigorous development and operational transformation that spans from chip to utility. Complimented by our clean energy contracts, strategic investments, and steadfast commitment to water stewardship, we are actively paving the way for responsible, sustainable advancement.
-
Sridhar Lakshmanamurthy liked thisSridhar Lakshmanamurthy liked thisCatching the silicon architecture team in a deep, philosophical debate about... whether that's an espresso or a cortado. The architecture looks solid, but the coffee situation requires additional meetings. ☕️😄 Want to help us design the future of silicon (and maybe join these rooftop coffee breaks)? We’re hiring! Check out our open roles or reach out directly. 🚀 #SiliconLife #GoogleHardware #Hiring Rafi T.Shuki YabboAviel NahoumNir GerberYiftach BenjaminiYoav BabajaniOphir Edlis
-
Sridhar Lakshmanamurthy liked thisSridhar Lakshmanamurthy liked thisFor years, AI has largely been a cloud story. The biggest models needed the compute, memory, and scale that only the data center could provide. That’s starting to change. As models become smaller and devices become more capable, more intelligence can move closer to the user. And that raises a more interesting question than edge versus cloud: What should run where, and why? Cost, privacy, latency, connectivity, power, and performance all shape that answer. In many cases, the best architecture will use both the edge and the cloud, with workloads moving between them based on what each environment does best. In this latest article of mine, I explore why the next phase of AI may be less about where intelligence lives and more about how intelligently we design the systems around it. #AI #EdgeAI #CloudComputing #Semiconductors #MediaTek
-
Sridhar Lakshmanamurthy liked thisSridhar Lakshmanamurthy liked thisAir cooling in the data center has officially hit a physical wall. ASHRAE standards and facility benchmarks put the practical ceiling for air cooling around 30kW–35kW per rack. When frontier AI clusters push past 100kW, the physics simply break down. You can't force enough air through a standard rack to extract that heat without wasting massive power on fans and triggering thermal throttling. At that density, direct-to-chip liquid cooling isn't a specialized experiment. It's baseline engineering. When your thermal architecture fails, your accelerators sit idle, unit economics fall apart, and training runs stretch out by weeks. That is why our teams focused on direct-to-chip liquid cooling across our TPU fleet starting back in 2018. Co-designing thermal management into both the silicon and the facility fabric is what makes exascale clusters viable. How is your team handling the 30kW limit on air cooling in your cluster deployments this year? https://lnkd.in/gVaQ3Y-7 #Google #TPU #DataCenters #AIInfraInside the Ironwood TPU codesigned AI stack | Google Cloud BlogInside the Ironwood TPU codesigned AI stack | Google Cloud Blog
Licenses & Certifications
Languages
-
Hindi
-
-
Telugu
-
-
Tamil
-
Other similar profiles
Explore more posts
-
Horace Chan
I ask the frontier and open source models to write a simple ASCII stock chart program that fetches data from yahoo finance. Opus 5 and GLM-5.2 are most expensive, spent over $10 writing this simple app GPT-5.6 Sol and Deepseek v4 Pro is the cheapest, costing less than $0.5 Gemini 3.7 flash is the only model that can't even finish this task. 🤔 Pick your model wisely, or you are really flushing money down the toilet
1 Comment -
Griner Solutions LLP
Accelerate your data center and AI performance with our latest stock! We have a new inventory of top-tier components from industry leaders like NVIDIA (Mellanox), Samsung, and Intel, available now for immediate delivery. Our stock includes essential hardware for building and scaling high-performance computing (HPC), AI, and enterprise data center environments. High-Speed Networking: (2x) MCX75310AAS-NEAT :NVIDIA ConnectX-7 400Gb/s Single-Port OSFP VPI Adapter Cards, PCIe 5.0 (2x) MCX755106AS-HEAT :NVIDIA ConnectX-7 200Gb/s Dual-Port QSFP112 VPI Adapter Cards, PCIe 5.0 Next-Gen Server Memory: M321R8GA0PB0-CWMXJ - Samsung 64GB 2Rx4 DDR5 RDIMM 5600 MT/s M393A4K40EB3-CWE - Samsung 32GB 2Rx4 DDR4 RDIMM 3200 MT/s M321R4GA3PB0-CWMKJ - Samsung 32GB 2Rx4 DDR5 RDIMM 5600 MT/s (similar spec) High-Performance Processing: CL8064701483802,SR17L - Intel Core i7-9700KF 9th Gen 8-core Processor CAN Bus Protection: PESD2CANFD24V-TR - Nexperia ESD and Surge Protection for CAN FD bus applications Contact Us: Interested in these parts or other components for your project? Send us a DM or visit our profile to learn more. Our inventory is ready to ship. #HPC #DataCenter #AI #Networking #Servers #Hardware #NVIDIA #Intel #Samsung #Inventory #TechUpgrade #ConnectX #DDR5 #EnterpriseTech
-
Wei-shyong Lee
👍 Good Reading this Industry Insight Article: Cost Dynamics Across Pre-Silicon & Post-Silicon in Chip Productization 👉 Read here (chetanpatil.in) 💡💡 My personal take away : Silicon complexity, Customer Expectations, and Why Validation is the last safety net with Shift Left Strategy !! What Post-Silicon Validation Is • Runs on prototype boards with real OS, apps, and stress workloads. • Focuses on system-level correctness (timing, power, interactions). • Complements manufacturing test, which only screens defects. Why It Matters 1. Cost to Fix • Pre-silicon fix: simulation rerun • Post-silicon fix: lab debug, re-spin, regression • Post-launch fix: recalls, lost market share, reputation damage 2. Industry examples: chip errata, recalls, or delayed launches due to missed corner cases.
1 Comment
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content