Integrated Matrix Extension

• Task Group Info
  • Group page
  • Mailing list
• Meeting Details
  • Official RISC-V Technical Calendar
  • Date: Odd weeks on Mon.
  • Time: 08:00 (PDT), 11:00 (EDT), 12:00 (BRT), 16:00 (CET), 23:00 (CST/TST), 24:00 (JST).
  • Zoom link
  • Meeting ID: 933 9715 7767
  • Passcode: 051351
• Disclaimer
  • Slides
  • Video
• Acting Chairs
  • Guido Araujo (Acting Chair, UNICAMP)
  • Jose Moreira (Acting Vice-chair, IBM)
• Meetings
  • Next meeting: 2024/08/09
  • Previous meeting: 2024/07/29
  • All Minutes
  • All Slides
  • Videos
• Roadmap
• Working groups
• Articles
• Presentations
  • Architectures
  • Memory Analysis
  • uArch
  • Workloads
• Notebooks and code
  • [Andes and Unicamp] Computational intensity notebook
  • [Unicamp] Burst analysis notebook
  • [Andes] Cache and DRAM evaluation tool
  • [Unicamp] Revisiting CI under a datatype glass
• Workloads and Benchmarking
  • [Alibaba] Analysis of matrix and vector instructions of ML Models
  • [Unicamp] Analysis of vector and matrix transfers in Polybench (HPC benchmark)
  • [Unicamp] ConvBench
    • Repository
    • Slides
    • Video
  • [Eldorado and Unicamp] IBM POWER10 MMA Vector-Matrix Interaction
    • Thanks to Jose Moreira!
    • Description
    • Data
• Market Analysis
  • Companies Need On-Premise HPC – and for More Than AI, Too
  • Hyperion Research Analysis
• Mathematical Analysis
  • Lemmata
    • k = 1 maximizes computational intensity under architected state constraints
    • Computational intensity is maximized for m = n
  • Corollaries
    • When vector registers are used as accumulators maximum computational intensity is ½ * (sqrt(RV + 1) – 1).
    • When vector registers are used only as operands and accumulators are newly designed states, computational intensity is always larger and is given by RV/4.