Optimal weight splitting improves accuracy in noisy memory AI chips
Closed-Form Optimal Weight Splitting for Matrix-Vector Multiply on Noisy In-Memory Accelerators
Emerging Technologies
Summary
Performing neural network calculations inside memory chips saves energy and time, but these chips are noisy and can’t store exact numbers well. The authors focus on splitting each number across multiple stored values to balance noise and precision. They found a formula to choose the best way to split these values based on measuring the chip’s noise, improving accuracy over common methods. This helps make energy-efficient AI hardware work better.
What this means in practice
- •For hardware engineers: Adjust weight encoding in analog AI chips to improve accuracy by choosing optimal base splitting guided by device noise.
- •For ai accelerator designers: Design analog in-memory neural network chips with a closed-form formula to balance noise and precision in weight storage.
Authors
Mehmet Aktukmak, David A. Roberts
Abstract
Analog in-memory computing performs the matrix-vector multiplications of neural networks directly inside the memory array, which saves time and energy. But each memory cell stores only a few levels and is noisy, so one cell cannot hold a weight precisely. The standard fix is to split each weight across several cells, like digits of a number. The base of this split decides the trade-off: a large base gives fine resolution but amplifies the noise, and a base near one averages the noise but wastes resolution. The best-known prior analysis considers only the noise, and so it always picks a base of one. We propose a principled method that also accounts for the limited number of levels per cell. It gives the optimal base in closed form, from a single measurement of the device noise. On a device-physics simulator, the optimal split is 3.3 dB more accurate than the common base-16 split, and it beats equal splitting in end-application accuracy.