Benchmarking the AI Vector Accelerator (AVA) with CV32E40X using the CV-X-IF
Presented by Oana Lazar, Kunal Dalal, Adomas Lebedys & Yu Xia, University of Southampton
The AVA coprocessor implements a small subset of the RISC-V Vector extension, intended to accelerate AI inference applications. The point is to demonstrate that considerable speedup can be achieved using a minimal instruction set extension (in this case 8 instructions).
The AVA project has created a reference implementation using the Open Hardware Group’s CV32E40X core. This implementation is particularly notable for being the first use of the Open Hardware Group’s X-Interface, designed to make the creation of custom instruction set extensions as simple as possible. The benefit is measured using the TinyMLPerf benchmarking suite.
In this talk and discussion, the team will present the work to date, the challenges they faced, and the results they have achieved as part of this Master’s project sponsored by Embecosm.
Oana Lazar is in her fifth and final year studying MEng Electronic Engineering at the University of Southampton. During her fourth year, she completed a 12-month UKESF industrial placement with Tessent Embedded Analytics (formerly UltraSoC and now a Siemens Company) working on the Secure-CAV project. Her interests include embedded software for security applications, cybersecurity, and machine learning. She was recently named the 2021 UKESF Scholar of the Year.
Yu Xia is a final year MEng Electronic Engineering student at the University of Southampton studying modules focused on computer systems and VLSI design. Yu’s interests are within RTL design and verification for RISC computer architectures, previously interning
at Arm as part of his UKESF scholarship scheme. He will be joining Arm in September as a graduate systems engineer.
Kunal Dalal is in the final year of his MEng in Electronic Engineering at the University of Southampton. Alongside his part-time role in Application Engineering at Arm as an undergraduate, he is undertaking research with members of the POETS project at
the University of Southampton. Kunal’s interests lie in the RTL design of different accelerator microarchitectures, and the deployment of heterogeneous architectures in general.
Adomas Lebedys is in the final year of his MEng degree in Electrical and Electronic Engineering at the University of Southampton, with interests in systems programming, computer architecture, and machine learning. Outside his course, he leads the design of
automotive electronics and embedded software for the Southampton University Formula Student Team’s electric vehicle.