Filters and Redundancies: An Exploration of Novel Coherent Noise Filters for High Energy Physics
Felipe Costa, Nicolas Guimarães, Guilherme Milani, Bruno Sanches, Irakli Mandjavidze, Damien Neyret, Wilhelmus Van Noije
TL;DR
The paper evaluates radiation-tolerant coherent noise filtering for the SALSA front-end in the ePIC detector by comparing two median-finding algorithms, BWMF and CSMF, under three redundancy schemes (simple TMR, full TMR, TTMR) in a 65 nm process. BWMF provides low area and power with simple TMR but incurs high power under full TMR, while CSMF is more area-intensive but offers ultra-low latency. TTMR, applied to CSMF, delivers reliable fault tolerance with substantially reduced power and a modest area increase, outperforming full TMR for this design. The work demonstrates that a TTMR-enabled CSMF is a favorable, radiation-tolerant choice for CMN subtraction in SALSA. Overall, the study informs design choices for robust, low-power median-finding in high-energy physics readouts, balancing fault tolerance, area, and power constraints.
Abstract
This work presents radiation-tolerant implementations for the SALSA front-end readout ASIC through redundancy methods applied to two median-finding algorithms designed for coherent noise suppression. Bit-wise Median Finder (BWMF) and Combinatorial Sum Median Finder (CSMF) were implemented in TSMC \SI{65}{\nano\meter} and evaluated in terms of area, power, and latency. Three redundancy techniques were applied in this work to compare their impact: simple TMR, full TMR, and temporal TMR (TTMR). The simple and full TMR approach was applied in both algorithms to establish comparisons and TTMR was applied to CSMF as an improvement. The results indicate that the BWMF achieves efficient performance in terms of area and power under the simple TMR scheme, but exhibits significantly higher power consumption when using the more robust full TMR approach. The TTMR technique, in turn, offers reliable fault tolerance while maintaining a feasible balance between area and power.
