Real In-Memory Processing
Talkat Algorithmic Foundations for Emerging Computing Technologies Boot Camp 2025 · 2025-09-08..09-12·
summary
Claims a working memristive memory processing unit (mMPU (memristive MPU)) that implements MAGIC NOR stateful logic across 512×512 crossbars and exposes three parallelisms (row/column, partition, crossbar) to run millions of logic ops in parallel. Implements AritPIM for fixed/floating add/mul/div (bit-serial or bit-parallel) and reports up to ~700× throughput gain for addition and ~150× for division versus a GPU on 64M-element vectors. Adds diagonal ECC with a separate check memory (≈20% latency cost, orders-of-magnitude reliability gain) and a Python stack (PyPIM) for automatic mapping. Relational DBs show ~5–10× speedups; large matrix multiply and DNNs do not.
AI-written summary · about
chapters (9)
- 0:00Introduction to mMPU
- 2:40Circuit and device level
- 13:36Architecture abstraction
- 18:41Arithmetic algorithms
- 32:44Reliability in mMPU
- 39:30Programming and integration
- 47:32Database applications
- 52:18Neural network evaluation
- 58:09Conclusion and Q&A
related talks
01h 11mMachine Learning Acceleration in an Embedded World
19mTurbocharge ANNS on Real Processing-in-Memory by Enabling Fine-Grained Per-PIM...
58mKeynote: Enabling the AI Revolution: Next-Generation Memory Solutions for AI Computing(…)
24mMemristor-based neuroelectronics for artificial and hybrid intelligence
01h 57mIntroduction to Neuromorphic Computing
43mKeeping Moore's law alive: Neuromorphic computing
22mAdvancing AI Computing With Optimized Matrix Multiplication Techniques With RISC-V CPU
27mPyTorch Symmetric Memory: A New Programming Paradigm for Distributed AI