Ram Velega (Broadcom) at OCP "The things that have confined us today that the scale-up domain is less than 100 are going to be changed very soon" (7:55 in buff.ly/9ioPH73) I see this as a call to action for systems and software research. UE 1.0 techniques such as LLR and CBFC play a crucial role!

Ilya Sutskever "Three lines of math can prove all of supervised learning. That's nice" (4:33) "I have not seen an exposition of unsupervised learning that I found satisfying" (7:50) Its optimization objective has little relation to the actual objective you care about Watch: buff.ly/Bnvazym

Can we build an #AI #Climate Scientist? Asked at the ADIA Lab Symposium in Abu Dhabi last week - now online at buff.ly/6igSeyg :-). Much work to be done - this is outlining some directions of indicative results with a lot of potential to accelerate AI for Science.

Collaborator and friend Dan Alistarh talks at ETH about using the new NvFP4 and MXFP4 block formats for inference. Some going from "terrible" accuracy to acceptable using micro rotations to smoothen outliers in blocks. arxiv.org/abs/2509.23202 Great collaboration and cool stuff

Keren Bergman at the 2nd EFCL Workshop: "Huawei combines 3x less performant GPUs with a photonic scale-up network to build higher performance PODs than Nvidia." Nvidia moving towards CPO 😀. Optics everywhere in scale-out. Nice overview of optical networking at all distances.

I was very honored to meet Carnegie Mellon University's President, Dean of the School of CS, and its famous founder Raj Reddy to present a lecture named after him. I tremendously enjoyed speaking with young students and faculty and the evening with CMU's leadership. Thanks for the invitation Raj!

MIT's Sandy Pentland at ADIA Lab symposium: "Modern companies need to structure their incentives to coordiate teams instead of using strict and siloed hierarchies." Enable team leaders to do "what they think is right" instead of inefficient political discussions with leadership.

🎉 Uno accepted to SC25! Unified congestion control + reliable connectivity for intra- & inter-DC traffic to enable inter-DC AI training. 📄Paper: arxiv.org/abs/2510.15802 💻Code: github.com/spcl/Uno_SC25 🤝Collaboration with Microsoft #SC25 #AI #SPCL @thoefler.bsky.social @csateth.bsky.social

MIT's Sandy Pentland at the ADIA Lab symposium 3rd day on #AI in Finance: "More social traders minimize risk using collective (social?) intelligence." This applies to many fields! Do we need social #AI? Corollary: We need more #HPC compute for such social agents :-).

One highlight of the ADIA Lab Symposium was Nobel Laureate Chu's talk towards Net-Zero emissions. "China follows the US textbook from 100 years ago, when the US took products inventend in Europe, such as cars, industrialized and improved them. Now China takes things invented in the west..."

Just arrived at the ADIA Lab symposium in Abu Dhabi to listen to Horst Simon's introduction and Bjorn Stevens' keynote on how to compute the future climate! Featuring our Gordon Bell finalists 🌍🚀 Looking forward to speculating about how to create an #AI climate scientist 😀.

Microsoft's Ultra Ethernet tutorial is now available on youtube 🎥! Saurabh Dighe gives a brilliant motivation for Microsoft's "AI first datacenters" 🤖 followed by Abdul Kabbani and myself explaining the technical details and innovations of Ultra Ethernet supporting this goal 🫡. buff.ly/IPN46ZR

Next was an intriguing talk by @thoefler.bsky.social on computational architectures for more efficiently training large models at @scsatcmu.bsky.social www.youtube.com/watch?v=LnXp... (4/8)

I'm excited to discuss whether we can build an "AI Climate Scientist" in my talk at the ADIA Lab Symposium 2025 🌎 Join us in Abu Dhabi or online from October 27–29! Register here: buff.ly/gCP2K1z #ADIALabSymposium2025

I was shocked to see the first two people in my "masterclass" 🎓 on #AI networking with Ultra Ethernet at #HLF25: David Patterson and Bob Metcalfe 😅! Both Turing award winners - Bob being one of the inventors of Ethernet 🥹. Was great fun also with many enthusiastic students and great discussions 🚀.

I'm honored to present the Raj Reddy Artificial Intelligence lecture at Carnegie Mellon University this Thursday. Join us in person at 5pm in Rashid Auditorium, Gates Hillman 4401! buff.ly/1kj9LtN This lecture series is honoring the Turing award winner's work in AI.

Thanks to The New York Times for featuring the ACM A.M. Turing Prize! Indeed, beyond the Nobel Prizes, "there are plenty of other prizes scientists and mathematicians can compete for." Learn more: buff.ly/NAhXVmy

The first ADIA Lab transactions edited by Horst Simon arrive at my desk 📖. An exciting mix of different science areas under the umbrella of advanced scientific computing and #AI 🎯. Congratulations to Horst and all co-authors 👏. Onward 🚀!

Listen to 2024 ACM Prize in Computing recipient Torsten Hoefler share how he tries to bridge academia and industry for societal impact . Full episode here or wherever you get your podcast: learning.acm.org/bytecast/ep7... #ACMByteCast #computing

Cool and impactful project led by the MPI-M: coupled simulation of the full Earth system in 1.25km resolution. Unprecedented complexity, fidelity, and performance at 82.5 simulated days per day. Accelerated by DaCe on GH200 - next generation #HPC for #climate science is here!

I met many of my heroes at my first Heidelberg Laureate Forum #HLF25! Exciting discussions with 28 other laureates of the five highest awards in CS and Math and young researchers! Watch my Spark talk at: buff.ly/SO6ntUb Thanks to the HLF Foundation and Klaus Tschira Stiftung.

Happy 100th Birthday to Seymour Cray! 🎂 We took your relentless pursuit of speed and parallel processing, and built on your ideas to start the #AI revolution. Every neural network owes a debt to your chips. Thanks for the gigahertz, Seymour! We're putting them to good use - your #LLM friend. 😉 #AI

Listen to 2024 ACM Prize in Computing recipient Torsten Hoefler discuss the power of High Performance Computing. Full episode here or wherever you get your podcast: learning.acm.org/bytecast/ep7... #ACMByteCast #computing

Apertus, the Swiss Fully-Open-Data model downloaded more than 379k times - ~13 downloads per minute of up to 140+ GiB! Trending for some weeks among the world's top LLMs. The techreport is now on arXiv: buff.ly/bvSioQH

Bill Gropp speaks at the Modeling to Learning with #HPC #ICERM workshop about MPI performance, emphasizing the characteristic rendezvous "little blip" where bandwidth is lost in transition. This will completely go away to a smooth transition with Deferrable Sends in Ultra Ethernet! buff.ly/0bEBb6y

Microsoft's first of many (connected!) Fairwater GB200 Supercomputers coming online in Wisconsin. 10x faster than the top of the #top500 list powered by an innovative network after years of planning and design. Proud to have contributed :-). Watch out for more info appearing.

Wonderful ICERM meeting on #HPC and #AI honoring Bill Gropp from NCSA. After crossing five US states by car (NJ, NY, CT, RI, MA) I met many old and new friends. Thanks to David Keyes to take the lead of this in a multi-year planning effort! buff.ly/fu8h4mB

Was great to attend the Intel Academic Security workshop in Hillsboro, OR to talk about Marcin's work on performance aspects of Confidential HPC and AI systems as well as giving a keynote on UE from a security perspective. TSS rules - thanks to Eric Spada's leadership! Read more: buff.ly/A9jczpC

📢REPS accepted at EuroSys 2026!🚀 A per-packet load balancer for out-of-order transports. It caches high-performing paths and reroutes away from failures. It requires no switch changes and uses only ~25B per flow. 👉https://arxiv.org/pdf/2407.21625 #HPC @thoefler.bsky.social @csateth.bsky.social

New #ISC25 podcast with Kevin Jackson on ISC highlights on connecting everything, #AI vs. #HPC? No! HPC is the Engine for AI and AI is HPC. Precision boosting, Ultra Ethernet, Reasoning Language Models, AI applications, and Supercomputing as driver for future societies!

The paper is really good. It makes me wish my new job had me working on more fabric-related stuff. The networking inside storage systems is not very exciting.

Happening right now!🚀 Afif Boudaoud from @spcl-eth.bsky.social is presenting today in the AI Models & Approaches session at IEEE Cluster 2025 in Edinburgh! Check out our paper on DaCe AD👉 arxiv.org/abs/2509.02197 #HPC @thoefler.bsky.social @csateth.bsky.social

📢 Our paper Demystifying Chains, Trees, and Graphs of Thoughts just got accepted by the journal IEEE Transactions on Pattern Analysis and Machine Intelligence!🚀 👉https://ieeexplore.ieee.org/document/11123142 👉https://arxiv.org/abs/2401.14295 #HPC @thoefler.bsky.social @csateth.bsky.social

Learn about Ultra Ethernet 1.0 from some of the authors. Abdul Kabbani from Microsoft and I will be speaking about details of the 1.0 specification on Mon Sep. 8th 10am PST. Join in-person at Microsoft Redmond building 99/1915 (open to public) or online buff.ly/vWIGYls More info: buff.ly/ijiW5so

Congrats to Switzerland 🇨🇭 to its first sovereign #LLM #AI #chatbot. Apertus is a fully open (data) model trained on CSCS' 10,752 GH200 superchips #HPC. Like a proper Swiss - multi-lingual, ethical, neutral, and open (direct) 😅! Try: publicai.co Download: buff.ly/sbZJOyo Study: buff.ly/dHfcjkL

In this paper is presented an ab-initio transistor simulation of unprecedented scale including electron-electron interactions within the self-consistent GW approximation, carefully optimized to take advantage of the Alps and Frontier supercomputers. #HPC arxiv.org/pdf/2508.19138

The three tiers of openness for AI models: open weights, open source, and open data. They determine scientific reproducibility, analyzability, and integrity of released models. #AI for #Science buff.ly/MfZwKM3

3200 published papers this year 🤯

Understanding Data Movement in Tightly Coupled Heterogeneous Systems: A Case Study with the Grace Hopper Superchip”. We explore the performance of unified memory, intra-/inter-node operations on #NVIDIA GH200 in the Alps supercomputer, shedding light on #HPC and #AI tradeoffs. 🔬📊 buff.ly/ATuIldr

It can be daunting to read all 562 pages of Ultra Ethernet 1.0 for next-generation #AI and #HPC networking. Thus, we prepared an author's cut paper about the key principles, architecture, and innovation. buff.ly/tlW4wYD I will also give a distilled short talk on this at hoti.org next week!

Congratulations not only to Singapore at it's 60th birtday but also colleague and friend Satoshi Matsuoka for his recognition with the Public Service Medal (Friends of Singapore) by the President of the Republic of Singapore! I'm looking forward to collaborating and contributing more - maju!

Fond memories of the Ultra Ethernet membership summit in April. This time on an alpine route onto a real summit in 🇨🇭 ⛰️ :-). Climbing to new heights with the release of the UE 1.0 spec and full steam ahead enabling fundamental features for next-generation #AI and #HPC networking!

What is AI for Science? A technique that combines compressible and uncompressible knowledge as represented by methods in modeling & simulation and data science using high-performance computing to accelerate science and human progress at large. #HPC for #AI buff.ly/bcc8a4C

Big congrats to Andrei Ivanov for passing his PhD exam! 🥳 SPCL's 12th PhD student 🎓 released into the wild to make the world a better place 🌍. Many know his first work "Data Movement is All you Need" (buff.ly/BRAgvpN) making him an expert in #HPC and #AI optimization. 💪

Salishan Random Access talk: "Reasoning Language Models Will Solve All Our Problems (with the right machines)!" An intuitive overview of how these models, given enough compute power, can become scientists to power the future of all of us. 🧠💡☁️🔬 #AI #HPC buff.ly/cjb66xX

Scaling #LLM training across multiple datacenters is becoming increasingly crucial, but higher latency & limited bandwidth introduce major performance bottlenecks. Ales Kubicek presented a solution to this at ATC’25!🚀👇 www.usenix.org/conference/a... #HPC @thoefler.bsky.social @csateth.bsky.social

Noam Brown (OpenAI): "There is a tradeoff (in Hex): 10x in training time compute is equivalent to 15x test time compute" - interesting tradeoffs in reasoning tasks. Will all depend on the final use-case and deployment! buff.ly/ffdzLZx

With Tomahawk Ultra, Broadcom asks who needs UALink when there's Ethernet?

🚀 Impressive ISCA’25 paper dissecting DeepSeek’s 🧠 compute stack: buff.ly/fYYlCnN 💡 FP8 precision, MLA, MoEs, Multi-plane Fat Trees w endpoint routing via NCCL’s PXN! 📐 The topology? Equivalent to a 1D HammingMesh for ~2k nodes. 2D scales further (shameless plug): buff.ly/YUF4Sqm #HPC #AI #ISCA25

🖥️ Excited for ROSS at #SC25—where operating and runtime systems for HPC take center stage. If you’re working on bold innovations or clever solutions, we’d love to see your paper submission. Join us in Dallas to shape what’s next: buff.ly/IRxleqE ⚙️ #HPC #AI #OperatingSystems