guest ::
| Home > Publications database > Fast Multi-Tau Correlators on FPGA with Context Switching From and to High- Bandwidth Memory |
| Home > Publications database > Fast Multi-Tau Correlators on FPGA with Context Switching From and to High- Bandwidth Memory |
| Contribution to a conference proceedings/Contribution to a book | PUBDB-2026-00685 |
; ;
2025
IEEE
ISBN: 979-8-3315-5752-2
This record in other databases:
Please use a persistent id in citations: doi:10.1109/ICFPT67023.2025.00027
Abstract: Various methods to measure the dynamic behavior of particles require the calculation of autocorrelation functions. For this purpose, fast multi-tau correlators have been developed in dedicated hardware, in software, and on FPGAs. However, for methods such as X-ray Photon Correlation Spectroscopy (XPCS), which requires to calculate the autocorrelation function independently for hundreds of thousands to millions of pixels from high-resolution detectors, current approaches rely on offline processing after data acquisition. Moreover, the internal pipeline state of so many independent correlators is far too large to keep it on-chip. In this work, we propose a design approach on FPGAs, where pipeline contexts are stored in off-chip HBM memory. Each compute unit iteratively loads the state for a single pixel, processes a short time series for this pixel, and afterwards writes back the context in a dataflow pipeline. We have implemented the required compute kernels with Vitis HLS and analyze resulting designs on an Alveo U280 card. The design achieves the expected performance and for the first time provides sufficient throughput for current high-end detectors used in XPCS.
|
The record appears in these collections: |
PDF
PDF (PDFA)