A System-Level Evaluation of Q.ANT Native Processing Units in an HPC Environment
Authors/Creators
Description
This report evaluates Q.ANT’s Generation-1 and Generation-2 Native Processing Units (NPUs), photonic analog accelerators investigated within the DAQC project extension at the Leibniz Supercomputing Centre (LRZ). Benchmarks covering matrix operations, neural network training and inference, and image preprocessing assess performance, energy consumption, numerical accuracy, and suitability for high-performance computing (HPC).
Generation-2 achieves substantial improvements over Generation-1, including speedups exceeding 50× for matrix-matrix multiplication and 25× for ResNet-18 inference. The accelerator card energy consumption for matrix–matrix multiplication decreases by approximately a factor of six. Numerical accuracy supports the evaluated AI and image preprocessing tasks, although analog noise limits effective precision to approximately 6–7 bits.
Digital-analog conversion, data movement, electronic processing, and limited compiler support constrain their applicability. The results support further investigation, particularly for workloads tolerant of reduced precision, while identifying tighter optical-electronic integration and improved programmability as essential for broader HPC adoption.
Files
LRZ-Q.ANT-report-12NOV26.pdf
Additional details
Additional titles
- Subtitle (English)
- Photonic Analog Computing
Funding
Dates
- Issued
-
2026-11-12