Hardware-Aware Static Optimization of Hyperdimensional Computations
Explore a groundbreaking video presentation from OOPSLA2 2023 on hardware-aware static optimization of hyperdimensional computations. Delve into the innovative Heim framework, which analytically derives minimum hypervector sizes for binary spatter code-based hyperdimensional computing. Learn how this approach guarantees optimized computations that meet accuracy targets while minimizing resource usage, even in the presence of hardware errors. Discover the framework's effectiveness in achieving higher accuracy, reduced hypervector sizes, and faster parametrization compared to dynamic tuning-based optimization across 25 benchmark data structures. Gain insights into the performance benefits of using analog CAMs and multiple-bit-per-cell ReRAM over conventional hardware while maintaining iso-accuracy.