We are immensely proud to announce that the HiCREST Laboratory will have a strong presence at the upcoming SC26 (Supercomputing 2026), the premier international conference for High Performance Computing, Networking, Storage, and Analysis. This year, our group has achieved an outstanding milestone with three research papers accepted into the main technical program.
These papers span the cutting-edge domains of quantum computing reliability, hardware resilience for AI accelerators, and the scaling behavior of massive AI training workloads.
Here is a closer look at our accepted works:
Characterizing the Scalability and Performance of Large-Scale AI Training Under Multi-Tenancy
Authors & Collaboration: Work by Jacopo Raffi, Thomas Pasquali, Lorenzo Piarulli, Filippo Spiga, Marco Faltelli, Andreas Herten, Domenico Siracusa, Daniele De Sensi, and Flavio Vella, in collaboration with Sapienza Università di Roma, ENEA, Forschungszentrum Jülich, and NVIDIA.
The Research: This paper provides a systematic, large-scale picture of how concurrent AI training jobs interfere with one another on real-world supercomputers. The team tested synthetic proxy models on up to 2,620 GPUs across some of the world’s most powerful architectures, including Alps, Leonardo, LUMI, JUPITER, NVL72 GB300, and DGX A100.
TETRIS-Q: Tiling-based Effective Transient-fault Reduction with Interleaved Superconducting Qubits
Authors: Led by Marzio Vallero (who recently completed his Ph.D. within our group), with the support of Gioele Casagranda and Flavio Vella.
The Research: This work tackles quantum error correction (QEC) by proposing a novel cross-layer reliability solution. It innovatively combines strategic qubit placement, specialized barriers, and shuffling techniques to mitigate transient faults in superconducting quantum processors.
The Anatomy of a Misprediction: Tracking Terrestrial and Orbital Radiation-Induced Errors in Systolic Arrays
Collaboration: An incredible joint effort with the TARAN research group (Rafael Tonetto, Fernando Fernandes, Marcello Traiola, and Angeliki Kritikakou) and our close colleagues at The University of British Columbia (Abraham Chan and Karthik Pattabiraman).
The Research: To deeply understand and mitigate how radiation affects hardware-accelerated AI, this study combines physical beam experiments, RTL-level fault injections, and software-level fault injections. The results provide crucial insights into how terrestrial and orbital radiation trigger errors in systolic arrays.
These accepted works reflect the dedication of our researchers and the strength of our international and industrial partnerships. We look forward to presenting our findings and connecting with the global HPC community in Chicago this November!
