Torsten Hoefler 🇨🇭

The future of #DeepLearning is sparse! See our overview of the field and upcoming opportunities for how to gain 10-100x performance to fuel the next #AI revolution. #HPC techniques will be key as large-scale training is #supercomputing. arxiv.org/abs/2102.00554 #MachineLearning

Registration to @IEEEHOTI 2025 is open: hoti.org/register.html @LightmatterCo, @CornelisHPC, @Cisco, Nubis Communications, Avicena Tech, @AristaNetworks,@UltraAccelLink, Celestial AI, @Meta, @giga_io, @Lenovo, @Dell, @ComputerSociety, @HPCwire #hoti #hoti25

Switzerland named the most innovative country for the 8th year in a row! ✨ The European Commission’s Innovation Scoreboard 2025 once again places Switzerland at the top. ℹ️: bit.ly/40XqSpy

You solidify understanding by teaching it to someone else.

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 youtube.com/watch?v=ltiaV3…

Thank you @ultraethernet ! It has been an unforgettable ride from the first idea to initial discussions among a small number of companies, followed by the founding of UEC, and just recently, the release of the UE 1.0 specification! Thanks to all the great colleagues. Onward!

Scaling #LLM training across multiple datacenters is becoming increasingly crucial, but higher latency and limited bandwidth introduce significant performance bottlenecks. Ales Kubicek presented a solution to this at ATC’25!🚀👇 usenix.org/conference/atc… #HPC @thoefler @CSatETH

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! youtube.com/watch?v=eaAonE…

🚀 Impressive ISCA’25 paper dissecting DeepSeek’s 🧠 compute stack: arxiv.org/abs/2505.09343 💡 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: arxiv.org/abs/2209.01346

Traditionally excellent venue to present your HPC-related system software work!

🖥️ 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: sites.google.com/view/rossworks… ⚙️ #HPC #AI

Japan switches back to nuclear, 14 years after Fukushima on.ft.com/4ewDtFO

Noam Brown (OpenAI): "I want to really emphasize if you look at just the raw NN of AlphaGo Zero, even through it was trained with MCTS, if you run at test time without MCTS, the performance is below top humans" Compute at inference is key for reasoning. youtube.com/watch?v=eaAonE…

Closing the ML Gap: Congrats to Stefano Schuppli for receiving the Best Paper Runner-Up Award at CUG 2025!🚀A proud moment for CSCS and a big step toward more accessible, efficient ML on HPC. Find out more👉 arxiv.org/abs/2507.01880 #HPC @thoefler @CSatETH @cscsch

Noam Brown (OpenAI) points out that availability of compute were the major limitation in getting to reasoning earlier. "People just underestimated how much [compute] would make [in poker]" - "I would have predicted 10x, not 100,000x". youtube.com/watch?v=Gr_eYX… (around 11:30)

"Swiss deep tech companies [...] created more than $100 billion worth of enterprise value." The result of sustained investment into top-class education, research, computational tools, and collaboration. Every dollar spent multiplies the Swiss economy. forbes.com/sites/davidpro…

ISC concluded :-). With a final attendee count of 3.585, the biggest ISC ever. Thanks for the opportunity to lead this year's journey - I immensely enjoyed working with Tanja, Nages, and the team! I passed the baton to Rosa Badia. Onward! isc-hpc.com/isc-2025-concl…

One big benefit of waking up at 4am, jetlagged, is the early morning run! Getting ready to talk about how #HPC drives the future of #AI #LLM and RLM architectures at the IEEE Cloud Summit in Washington, DC. How can #XaaS Accelerated Containers help? arxiv.org/abs/2401.04552

Ultra Ethernet grows up quickly! 🚀 Our transport layer (UE Transport, UET) has officially been assigned UDP port 4793. Not only a beautiful prime number but also (++RoCE)++ 10 years later 😀 iana.org/assignments/se… Llooking forward to many #AI & #HPC implementations soon!

Nice evening with the Dieter Schwarz Stiftung and the next-generation #AI for #Science experts at ETH Zurich's new Heilbronn campus. Now on my way to the IEEE Cloud Summit to give the first keynote Thursday. I'm looking forward to see many of you there! ieeecloudsummit.org

-ETH Zurich confirms its place among the world’s best! 🇨🇭 According to the QS World University Rankings 2026, ETH Zurich ranks 7th in the world - for the third time in a row. Key factors behind its success? 🔬 Excellent reputation among scientists worldwide 📚 A high number of…