Hardware-Aware Static Optimization of Hyperdimensional Computations
summary
Heim is a hardware-aware static optimizer for BSC hyperdimensional computing that analytically derives the minimum hypervector size and distance thresholds and guarantees expected accuracy under device bit flips. It builds closed-form models of Hamming-distance means and variances, folds hardware error (e.g., 2.15% error rate) into the math, and outputs parameters in milliseconds. Compared to Monte‑Carlo tuning it finds parametrizations 30×–100k× faster, shrinks hypervectors 1.15×–7.14×, and meets 99% accuracy (median 99.2%–100.0%). Heim enables iso‑accuracy comparisons of emerging memories.
AI-written summary · about
chapters (9)
- 0:00The reliability challenge
- 1:32Hyperdimensional computing
- 2:40Introduction to Heim
- 3:31Encoding data structures
- 9:20Limitations of tuning
- 10:35Heim system architecture
- 12:36Technical implementation
- 14:38Evaluation and results
- 16:42Q&A and future outlook
related talks
21mRemoving Obstacles before Breaking Through the Memory Wall: A Close Look at HBM...
11mGeometric Methods for Machine Learning and Optimization
01h 21mThe Future of Hardware Technologies for Computing
15mHetu v2: A General and Scalable Deep Learning System with Hierarchical and Heterogeneous
05mBit-Blasting Probabilistic Programs
26mEfficient and Flexible Hyperparameter Optimization
56mReduce, Reuse, Recycle: Is this Mantra under attack with Proliferation of Machine(…)
48mComputational-Statistical Tradeoffs in Robust High-Dimensional Estimation