Verification Futures » VF2025 UK » Automating the design of bespoke AI accelerators for FPGAs

Automating the design of bespoke AI accelerators for FPGAs

Conference:	Verification Futures 2025 (click here to see full programme)
Speaker:	Alexander Montgomerie-Corcoran
Presentation Title:	Automating the design of bespoke AI accelerators for FPGAs
Abstract:	At the limits of power, performance and area, Edge AI's demand for low-latency, energy-efficient hardware challenges traditional inference capabilities. Whether it is Object Detection, Image Segmentation or Super Resolution, edge devices each have unique requirements and constraints for the application they are deploying. GPUs and NPUs struggle with inefficiencies in any particular one due to their generalized "one-size-fits-all" design approach. On the other hand, the versatility of FPGAs allow for highly specialized and therefore efficient hardware designs. To realize the potential of FPGAs for AI applications, Heronic has created a toolflow for designing bespoke AI accelerators for FPGA systems. This talk will cover how engineers can leverage the benefits of Heronic’s approach to accelerator design for FPGA devices.
Speaker Bio:	Alexander Montgomerie-Corcoran is the Co-Founder and CEO of Heronic Technologies. Prior to starting the company, Alex completed his PhD at Imperial College London, researching methods for accelerating AI on FPGA devices. His research led to several publications, state-of-the-art benchmark results on MLPerf Tiny, and 1st place in AMD’s OpenHW competition.
Key Points:	AI Acceleration Design Automationt Efficient Hardware Design

ML Verification Turns Convention on its Head

The verification of processor architectures designed for Machine Learning (ML) applications represent a departure from conventional techniques. Conventional constrained random testbenches, which focus on stimulus driving coverage, cannot scale for many ML algorithm realizations. ML architectures involve neural networks of processors that “learn” by manipulating coefficients across the network to match ideal outputs to a large quantity of input data. Furthermore, smart compiler technology is employed to leverage the many paths available in the network. An effective verification strategy can leverage planning algorithms that start with the desired output and optimize input values to achieve that output. Ensuring the paths that the compiler might trigger have all been tested, and that the test content can scale from individual processors to the entire network are critical challenges. Breker will share various approaches to this problem, developed through cooperation with three noted AI processor providers.

3 Key Points:

Current verification methodologies cannot scale to meet ML processor challenges

ML verification approach: consider desired outputs, optimize inputs to match

Test Suite Synthesis enable planning algorithm approach to target ML requirements

An Emulation Strategy for Artificial Intelligence Designs

The emergence of Artificial Intelligence is the “next big thing” and presents a unique opportunity for disruptive semiconductor development. End applications could range from ADAS, to 3D facial recognition, to voice and image processing, or to intelligent search. The SoCs for AI applications whether targeted for training or inference will have their own unique characteristics, but present quite common verification challenges that we will present in this session.

Supporting designs as big as 15 billion gates, Mentor’s Veloce Strato has unique virtualization capabilities that enable highly accurate pre-silicon execution of AI benchmarking applications like MLPerf. The Veloce Power App enables analysis of peak and average. We will cover how Veloce Strato and its supporting solutions are the best tool to help address the verification challenges of SoCs targeted for AI applications.

3 Key Points:

Deterministic solution for AI chips verification

Full virtual solution for HW/SW verification

TERAOPS/Watt assessement prior silicon availability

AI chips must get the math right

Most AI chips and hardware accelerators that power machine learning (ML) and deep learning (DL) applications include floating-point units (FPUs). Algorithms used in neural networks are often based on operations that use multiplication and addition of floating-point values. FPUs are difficult to implement. The IEEE 754 standard defines many corner-case scenarios and non-ordinary values. Even a minor rounding mistake could accumulate over many iterations and produce a large error. An FPU formal verification app compliant with IEEE-754 provides an efficient and rigorous solutions to FPU functional verification

3 Key Points:

Floating-point unit (FPU) for AI chips

FPU Formal Verification App

Compliance with IEEE-754

Name: Mike Bartley

Designation: Senior Vice President – VLSI Design

Title: Introduction

Biography:

Mike Bartley has a PhD in Mathematics from Bristol University, an MSc in Software Engineering, an MBA from the Open University and over 25 years of experience in software testing and hardware verification. He has built and managed state-of-the-art test and verification teams in a number of companies who still use the methodologies he established. Since founding TVS in 2008 he has grown the company to over 100 employees worldwide. Dr Bartley is Chair of both the Bristol branch of the British Computer Society and the West of England Bristol Local Enterprise Partnership (LEP). He has had over 50 articles and presentations published on the subjects of hardware verification, software testing and outsourcing.

Automating the design of bespoke AI accelerators for FPGAs

About Tessolve

Solutions

Tessolve Labs

Insights

Connect with us

ML Verification Turns Convention on its Head

An Emulation Strategy for Artificial Intelligence Designs

AI chips must get the math right

Please fill the form below

Please fill the form below

Please fill the form below

Please fill the form below

Please fill the form below

Please fill the form below

Please fill the form below