About
CTO at Physical Superintelligence. Previously at NVIDIA, IBM Research, and the Department of Energy's national laboratories.
At NVIDIA: led agentic GPU kernel development and optimization for frontier AI lab workloads, and drove performance evaluation strategy for CuTile, a next-generation GPU programming model.
At IBM Research: led IBM's enterprise generative AI coding assistant from research to production — 8-person team within a 60+ person cross-functional effort across IBM and Red Hat. Earlier led HPC systems work on Summit and Sierra (world's #1 and #2 supercomputers at launch), supporting large-scale scientific simulation workloads at DOE national laboratories. Founded NYU Courant's first graduate course on high-performance ML during this period, still taught today.
Before IBM: built distributed runtime systems at PNNL that spun off into Trovares (now RocketGraph).
Ph.D. in Computer Architecture, UPC Barcelona. IPDPS Best Paper Award (2012). 30+ publications, 15 patents.
Education
- Ph.D. in Computer Architecture — Polytechnic University of Catalunya (UPC), Barcelona
- M.Sc. in Computer Engineering — University of Rome Tor Vergata, Rome
- B.Sc. in Computer Engineering — Roma Tre University, Rome
Featured Work
CuTile Programming Model
AI-driven GPU kernel development and next-generation programming model for production workloads.
Watson Code Assistant
Led creation of the first generative model for IBM's AI-powered code assistant product.
Summit & Sierra Supercomputers
System software for the world's #1 and #2 fastest supercomputers (2018 TOP500).
Awards
- IBM Outstanding Technical Achievement Award (2023) — For leading the creation of the first generative model for Watson Code Assistant
- IBM Research Division Award (2022) — For contributions to AI for code
- HPCwire Editors' Choice Award (2018) — For Summit supercomputer
- PNNL Outstanding Performance Award (2015) — For contributions to extreme-scale computing
- IPDPS Best Paper Award (2012) — For research on scaling irregular applications on massively multithreaded systems
- HiPEAC Paper Award (2010) — For research on TLB misses in chip multiprocessors
In the Press
Physical Superintelligence (2026)
Announcing Physical Superintelligence, a public benefit corporation pointing superintelligence at the hardest problems in physics — producing new physics, at scale, discovered and commercialized safely and verifiably for broad public benefit.
- Physical Superintelligence — Dr. Alex Wissner-Gross, The Innermost Loop
CUDA Tile Programming Model (2025)
- NVIDIA Introduces CUDA 13.1 with CUDA Tile — InsideHPC
- NVIDIA CUDA 13.1 Powers Next-Gen GPU Programming — NVIDIA Developer Blog
AI for Code / Watson Code Assistant (2022–2023)
Summit & Sierra Supercomputers (2018)
- Two DOE Supercomputers Top List of World's Fastest — U.S. Department of Energy
- Sierra Honored as Top Supercomputing Achievement — LLNL / HPCwire
- Summit Supercomputer Is Already Making Its Mark on Science — HPCwire
Graph Analytics (2014–2019)
Selected Papers
For a full list, see Google Scholar.
AI and Machine Learning
- Towards Learning (Dis)-Similarity of Source Code from Program Contrasts
ACL 2022 - VELVET: a noVel Ensemble Learning approach to automatically locate VulnErable sTatements
SANER 2022 - D2A: A Dataset Built for AI-Based Vulnerability Detection Methods Using Differential Analysis
ICSE-SEIP 2021 - Probing Model Signal-Awareness via Prediction-Preserving Input Minimization
ESEC/FSE 2021 - Exploring Software Naturalness through Neural Language Models
ArXiv Preprint, 2020
High Performance Computing and Systems
- In-Memory Graph Databases for Web-Scale Data
IEEE Computer, 2015 - Scaling Semantic Graph Databases in Size and Performance
IEEE Micro, 2014 - Scaling Irregular Applications through Data Aggregation and Software Multithreading
IPDPS 2014 — Best Paper Award - Evaluating the Impact of TLB Misses on Future HPC Systems
IPDPS 2012 - A Quantitative Analysis of OS Noise
IPDPS 2011
Writing
Converting Compute Into Breakthroughs
For fifty years, software companies sold tools and left customers to turn them into outcomes. That arrangement has shaped the industry for decades — and it's now coming apart.
Read on LinkedIn →Tuning Flash Attention for Peak Performance in NVIDIA CUDA Tile
A deep dive into optimizing Flash Attention on Blackwell GPUs using cuTile, achieving 1.60-1.66x speedups through loop splitting, block remapping, and autotuning.
Read on NVIDIA Developer Blog →Generation Is Cheap. Verification Is Not.
Why complex software will survive the AI shock. The hard part was never writing the code — it was knowing whether the code was right.
Read on LinkedIn →The Physics of AI Systems: Four Fundamental Laws
AI systems have their own physics. Not metaphorical: actual constraints rooted in mathematics and the physical properties of silicon, light, and electrons.
Read on LinkedIn →We are the Master and the Slave
A parable about users and tools
Read on Substack →Intelligence Entails Ethics
The paperclip maximizer is a logical contradiction
Read on Substack →Get in Touch
Interested in collaborating on AI systems, GPU computing, or high-performance infrastructure? I'd love to hear from you.