Designing a Communication Library for Xilinx Versal Devices Using the Window-Based API
Open access
Author
Date
2023-08-20Type
- Bachelor Thesis
ETH Bibliography
yes
Altmetrics
Abstract
The Adaptive Compute Acceleration Platform (ACAP) developed by AMD/Xilinx is a novel architecture which combines three parts: An ARM-based CPU, an FPGA and a CGRA. The CGRA is implemented as a configurable grid of powerful vector processors called AI Engines (AIEs), whose communication patterns can be programmed using dataflow graphs.
Programming the AI Engines is difficult because it requires extensive knowledge of the underlying hardware. To the best of our knowledge, there are currently no open source libraries which add abstraction layers over the API provided by the manufacturers. Collective operations known from the Message Passing Interface (MPI) could simplify the programming experience, but only point to point communication and broadcasting are supported out of the box.
Using the VCK190 Evaluation Kit, which is part of the Versal AI Core Series, we build upon earlier exploratory work and extend it with a latency benchmark of inter-AIE communication for various distances. In the main part of the thesis, we adapt MPI’s Reduce function to this architecture as a particular example of a collective operation, and we implement our design. We perform various scientific benchmarks and develop a cost model to predict the cycle count of our implementation. Furthermore, we analyse several limitations which arise when scaling the implementation to larger data sizes and numbers of AIEs, which limits the usability of our prototype in real applications. Finally, we present theoretical considerations concerning the design of another collective operation called Gather. Show more
Permanent link
https://doi.org/10.3929/ethz-b-000671822Publication status
publishedPublisher
ETH ZurichSubject
AMD Xilinx Versal; Window Communication; Reduce; MPI; AI Engine Array; Versal Adaptive Compute Acceleration Platform; BenchmarkingOrganisational unit
03950 - Hoefler, Torsten / Hoefler, Torsten
More
Show all metadata
ETH Bibliography
yes
Altmetrics