EUVIMEDEUVIMED entdecken
Crossref

NeuSight: Predicting the Performance of Future AI on Emerging GPUs

Divya Mahajan, Seonho Lee, Junkyum Kim, Seokjin Go, Amar Phanishayee

IEEE Micro · 2026-09 · 46 (5) · 37-43

DOI: 10.1109/mm.2026.3704835 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Micro, 2026-09. Verfasst von Divya Mahajan und 4 weiteren Autor:innen. Die Arbeit nennt 3 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The rapid evolution of artificial intelligence (AI) makes it difficult for hardware and machine learning system designers to answer basic performance questions: How will current models perform on future GPUs? And how should models and hardware be co-designed? Answering these questions, today, typically requires scarce hardware, extensive benchmarking, high cost, and significant energy. We present NeuSight, a predictive framework that estimates training and inference performance on unseen GPUs. Rather than relying on regression-based or end-to-end black-box prediction, NeuSight decomposes GPU execution into tile-level components and constrains predictions using fundamental performance laws. A learned tile-level model captures locality, parallelism, and library-optimized behavior, then aggregates these effects to predict end-to-end latency. Beyond accuracy, NeuSight provides architectural insight by revealing when kernels are either compute or memory bound, where tensor cores and streaming multiprocessor parallelism are underutilized, and how bottlenecks shift across GPU generations. Across diverse workloads and GPU vendors, NeuSight generalizes accurately even on unknown models and hardware, substantially outperforming prior methods and enabling faster, lower-cost exploration of next-generation AI systems.

Zitieren

APA

Mahajan, D., Lee, S., Kim, J., Go, S., & Phanishayee, A. (2026). NeuSight: Predicting the Performance of Future AI on Emerging GPUs. IEEE Micro, 46(5), 37-43. https://doi.org/10.1109/mm.2026.3704835

Vancouver

Mahajan D, Lee S, Kim J, Go S, Phanishayee A. NeuSight: Predicting the Performance of Future AI on Emerging GPUs. IEEE Micro. 2026;46(5):37-43. doi:10.1109/mm.2026.3704835

BibTeX
@article{Mahajan2026,
  author = {Mahajan, Divya and Lee, Seonho and Kim, Junkyum and Go, Seokjin and Phanishayee, Amar},
  title = {NeuSight: Predicting the Performance of Future AI on Emerging GPUs},
  journal = {IEEE Micro},
  year = {2026},
  volume = {46},
  number = {5},
  pages = {37--43},
  doi = {10.1109/mm.2026.3704835}
}
Kostenlos starten →