We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.AR

Change to browse by:

cs

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Hardware Architecture

Title: When HLS Meets FPGA HBM: Benchmarking and Bandwidth Optimization

Abstract: With the recent release of High Bandwidth Memory (HBM) based FPGA boards, developers can now exploit unprecedented external memory bandwidth. This allows more memory-bounded applications to benefit from FPGA acceleration. However, we found that it is not easy to fully utilize the available bandwidth when developing some applications with high-level synthesis (HLS) tools. This is due to the limitation of existing HLS tools when accessing HBM board's large number of independent external memory channels. In this paper, we measure the performance of three recent representative HBM FPGA boards (Intel's Stratix 10 MX and Xilinx's Alveo U50/U280 boards) with microbenchmarks and analyze the HLS overhead. Next, we propose HLS-based optimization techniques to improve the effective bandwidth when a PE accesses multiple HBM channels or multiple PEs access an HBM channel. Our experiment demonstrates that the effective bandwidth improves by 2.4X-3.8X. We also provide a list of insights for future improvement of the HBM FPGA HLS design flow.
Subjects: Hardware Architecture (cs.AR)
Cite as: arXiv:2010.06075 [cs.AR]
  (or arXiv:2010.06075v1 [cs.AR] for this version)

Submission history

From: Young-Kyu Choi [view email]
[v1] Mon, 12 Oct 2020 23:23:18 GMT (489kb,D)

Link back to: arXiv, form interface, contact.