𝗕𝗶𝗻𝗲 𝗧𝗿𝗲𝗲𝘀: 𝗘𝗻𝗵𝗮𝗻𝗰𝗶𝗻𝗴 𝗖𝗼𝗹𝗹𝗲𝗰𝘁𝗶𝘃𝗲 𝗢𝗽𝗲𝗿𝗮𝘁𝗶𝗼𝗻𝘀 𝗯𝘆 𝗢𝗽𝘁𝗶𝗺𝗶𝘇𝗶𝗻𝗴 𝗖𝗼𝗺𝗺𝘂𝗻𝗶𝗰𝗮𝘁𝗶𝗼𝗻 𝗟𝗼𝗰𝗮𝗹𝗶𝘁𝘆 will be presented at SC Conference Series on Nov 20 in room 275!🚀🎉 📄 https://t.co/Lp2EQ5lHRM #HPC @thoefler @CSatETH #SC25

AI/HPC networks run real apps, not microbenchmarks. 🚀 Our 𝗦𝗖𝟮𝟱 Best Student Paper Candidate, 𝗔𝗧𝗟𝗔𝗛𝗦, traces real apps (like NCCL & MPI) to portable 𝗚𝗢𝗔𝗟 schedules for efficient simulation. 📄 Paper: https://t.co/Vu4EhlhgpM 💻 Code:

Happy to welcome Shriram Chandran, a new PhD student and look forward to our collaboration! 🚀 #HPC @thoefler @CSatETH

🎉 Uno accepted to SC25! Unified congestion control + reliable connectivity for intra- & inter-DC traffic to enable inter-DC AI training. 📄Paper: https://t.co/rgIWOSwuCo 💻Code: https://t.co/Bw0USVCnrK 🤝Collaboration with Microsoft #SC25 #AI #SPCL @thoefler @CSatETH

We are happy to welcome Pinxue Zhao, a new PhD student and look forward to our collaboration !🚀 #HPC @thoefler @CSatETH

SPCL researchers helped achieve a breakthrough in #climate modeling - running global simulations at 1.25 km resolution on Alps #HPC supercomputer 🚀This work makes decades-long runs feasible and is nominated for the Gordon Bell Prize for Climate Modeling. https://t.co/xcEPZcsFdR

Maciej presented at the GraphSys workshop at @euro_par in Dresden, the ACAT workshop in Hamburg, the Fast Machine Learning for Science Conference in Zurich, and a series of lectures at the Deep Learning Summer School at @AGH_Krakow in Krakow.

SPCL has been present at various events in the past weeks - Maciej gave keynotes and invited talks on the synergy between #graphs and #LLM, LLM reasoning, higher-order graph analytics, and other exciting topics in #HPC and #AI.

📢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://t.co/78PS8zTBG3 #HPC @thoefler @CSatETH

📢Our paper Psychologically Enhanced AI Agents is out ! We introduce MBTI-in-Thoughts, a framework for enhancing the effectiveness of LLM agents through psychologically grounded personality conditioning. Find out more: 👉 https://t.co/vBFDM3rCDB #HPC #AI @thoefler @CSatETH

📢 Our paper Demystifying Chains, Trees, and Graphs of Thoughts just got accepted by the journal IEEE Transactions on Pattern Analysis and Machine Intelligence!🚀 👉 https://t.co/6j0V3ZyWg4 👉 https://t.co/lDqJccNspG #HPC @MaciejBesta @thoefler @CSatETH

Happening right now!🚀Afif Boudaoud from @spcl_eth is presenting today in the AI Models & Approaches session at IEEE Cluster 2025 in Edinburgh! Check out our paper on DaCe AD👇 https://t.co/n3ouXfXV0N #HPC @thoefler @CSatETH

The analysis outcomes are synthesized in a set of insights that help to select the most beneficial GNN model in a given scenario, and a comprehensive list of challenges and opportunities for further research into more powerful HOGNNs. arXiv 👉 https://t.co/mC1YnpjmX7

To alleviate this, we first design an in-depth taxonomy and a blueprint for HOGNNs. This facilitates designing models that maximize performance. Then, we use our taxonomy to analyze and compare the available HOGNN models.

A plethora of HOGNN models have been introduced, coming with diverse neural architectures and notions of what the "higher-order" means. This richness makes it very challenging to appropriately analyze and compare HOGNN models, and to decide in what scenario to use specific ones.

Higher-order graph neural networks (#HOGNNs) are an important class of #GNNs that harness polyadic relations between vertices beyond plain edges. They are used to eliminate over-smoothing or over-squashing, enhance the prediction accuracy and improve the expressiveness of GNNs.

Paper 👉 https://t.co/88FtJ5X6Eg Dataset 👉 https://t.co/HQIN3UuHGq Code 👉

Excited to share our latest paper! We present a comprehensive survey and taxonomy of Filtered Approximate Nearest Neighbor Search (FANNS) algorithms, and we benchmark a selection of them on our novel *arxiv-for-fanns* dataset.

Cppless is open source and built on top of LLVM with less than 1k LoC changes. More details in the paper on serialization, C++ lambda extraction and cross-compilation. Paper 👉 https://t.co/BxREV1vZcf Code 👉 https://t.co/fOokcr5YMg Artifact 👉 https://t.co/AlSazEoRJ5

Our evaluation shows that C++ serverless functions can scale to 512 parallel workers with a double-digit millisecond overhead. On the example of ray tracing, we show a speedup of up 59x - from 60s → 1s execution time - with minimal cost increase.