Toward Lattice QCD On Billion Core Approximate Computers

TL;DR

This paper demonstrates that billion-core approximate computers can efficiently run simple lattice models with high accuracy, indicating potential for accelerating complex Lattice QCD simulations while reducing power and cost.

Contribution

It provides evidence that approximate hardware can run lattice models accurately and efficiently, paving the way for future LQCD computations on such systems.

Findings

Massively parallel approximate hardware achieves high speed and efficiency.

Similar results obtained with floating point and approximate representations.

Significant speed and power improvements over traditional CPU for large models.

Abstract

We present evidence of the feasibility of using billion core approximate computers to run simple U(1) sigma models, and discuss how the approach might be extended to Lattice Quantum Chromodynamics (LQCD) models. This work is motivated by the extreme time, power, and cost needed to run LQCD on current computing hardware. We show that, using massively parallel approximate hardware, at least some models can run with great speed and power efficiency without sacrificing accuracy. As a test of accuracy, a 32 x 32 x 32 U(1) sigma model yielded similar results using floating point and approximate representations for the spins. A 20 million point 3D model, run on a 34,000-core single-board prototype approximate computer, showed encouraging accuracy with a ~750 times improvement in speed and ~2500 times improvement in speed/watt compared to a traditional CPU. These results suggest there is value in future research to determine whether similar speed-ups and accuracies are possible running full LQCD on the compact billion-core approximate computing systems that are now practical.