# Talks

## 9 July 2020, Lecture

A field-programmable gate array (FPGA) is an integrated circuit in which the logic and interconnections can be defined by downloading a bitstream to its memory. Since their introduction in the 1980’s, these devices have dramatically increased in capacity and functionality, with the recent Xilinx ZU29DR FPGA integrating a 1.5 GHz Quad-core ARM Cortex-A53 processor; Dual-core ARM Cortex-R5; 16x 14-bit, 6.6 GSPS digital-to-analog converters; 16x 12-bit, 2.1 GSPS analog-to-digital converters; 4K DSP slices and 930K system logic gates on a single device. This presentation will first provide an overview of FPGA architectures and describe their Energy, Parallelism, Integration and Customisation (EPIC) benefits over other technologies such as microprocessors, digital signal processors and graphics processing units (GPUs). This will be followed with examples of interesting FPGA-based computing applications, including my own research in the Computer Engineering Lab at the University of Sydney.

## 1 June 2020, Lecture (Stephen Tridgell)

Real-time automatic modulation classification using RFSoC, 27th Reconfigurable Architectures Workshop (RAW)

The computational complexity of deep learning has led to research efforts to reduce the computation required. The use of low precision is particularly effective on FPGAs as they are not restricted to byte addressable operations. Very low precision activations and weights can have a significant impact on the accuracy however. This work demonstrates by exploiting throughput matching that higher precision on certain layers can be used to recover this accuracy. This is applied to the domain of automatic modulation classification for radio signals leveraging the RF capabilities offered by the Xilinx ZCU111 RFSoC platform. The implemented networks achieve high-speed real-time performance with a classification latency of ≈8µs, and an operational throughput of 488k classifications per second. On the open-source RadioML dataset, we demonstrate how to recover 4.3% in accuracy with the same hardware usage with our technique.

## 24 Mar 2020, Lecture

LUXOR: An FPGA Logic Cell Architecture for Efficient Compressor Tree Implementations, 28th ACM/SIGDA International Symposium on Field-Programmable Gate Arrays (FPGA), Monterey

We propose two tiers of modiications to FPGA logic cell architecture to deliver a variety of performance and utilization beneits with only minor area overheads. In the first tier, we augment existing commercial logic cell datapaths with a 6-input XOR gate in order to improve the expressiveness of each element, while maintaining backward compatibility. This new architecture is vendor-agnostic, and we refer to it as LUXOR. We also consider a secondary tier of vendor-speciic modifications to both Xilinx and Intel FPGAs, which we refer to as X-LUXOR+ and I-LUXOR+ respectively. We demonstrate that compressor tree synthesis using generalized parallel counters (GPCs) is further improved with the proposed modiications. Using both the Intel adaptive logic module and the Xilinx slice at the 65nm technology node for a comparative study, it is shown that the silicon area overhead is less than 0.5% for LUXOR and 5-6% for LUXOR+, while the delay increments are 1-6% and 3-9% respectively. We demonstrate that LUXOR can deliver an average reduction of 13-19% in logic utilization on micro-benchmarks from a variety of domains. BNN benchmarks benefit the most with an average reduction of 37-47% in logic utilization, which is due to the highly-efficient mapping of the XnorPopcount operation on our proposed LUXOR+ logic cells.

## 19 Feb 2020, Lecture

Low-precision CNN Inference and Training for Edge Applications on FPGAs, Xilinx Research Labs, San Jose

Much of the focus on FPGA-based implementation of convolutional neural networks (CNNs) has targeted server-based applications for which the primary benchmark set is Imagenet. In this talk, we discuss our research on FPGA-based inference and training for small convolutional neural networks, intended for edge applications such as radio frequency machine learning (RFML).

For inference, an open-source, ternary neural network generator which produces fully unrolled designs is presented. A VGG-style network implemented with this approach can achieve 90.9% accuracy for CIFAR10. On an Ultrascale+ VU9P, it achieves 122 k frames per second, with a latency of 29 μs. When applied to radio frequency modulation classification on a Xilinx ZCU111 board, we achieve 488 k classifications per second with a latency of 8 μs.

We also present a mixed-precision training accelerator, where most of the computations are processed using 8-bit block floating point. The work includes modifications to the stochastic weight averaging low-precision (SWALP) algorithm to achieve a 0.5% accuracy improvement for the MNIST and CIFAR10 benchmarks, bringing it within 0.1% of floating-point.

## 26th November 2019, Lecture

Multipliers for FPGA Machine Learning Applications, Imperial College

FPGA implementations of machine learning algorithms have received considerable interest, and multipliers are a key component for their efficient implementation. In this talk, four examples of our recent research efforts in customisation of FPGA architecture and designs using different approaches to multiplication be described. As the multipliers become more specialised, higher performance gains are possible.

First we describe a two-speed multiplier (TSM) which offers improved average clock rate over a conventional multiplier by carefully considering the delay of two critical paths in a serial-parallel modified radix-4 Booth multiplier. Secondly, our joint research with Fudan University on a flexible precision, run-time decomposable DSP architecture for CNN implementations which optimises Precision, Interconnect, and Reuse (PIR-DSP) is presented. The resulting design offers a 6× improvement in multiply-accumulate operations per DSP in the 9 × 9-bit case, 12× for 4 × 4 bits, and 24× for 2 × 2 bits. Thirdly, we present AddNet which shows that reconfigurable constant coefficient multipliers (RCCMs) offer an alternative to low precision arithmetic for saving area. Finally, we describe a ternary neural network generator which can generate fully unrolled designs. The implementation of a VGG-style network achieves 90.9% accuracy and 122 k frames per second, with a latency of only 29 μs, which is the fastest CNN inference implementation reported so far on an FPGA.

## 7th November 2019, Lecture

Computer Engineering Lab, Defence Innovation Network Industry Forum, University of Technology, Sydney

The Computer Engineering Lab at the University of Sydney focusses on how to use FPGAs to solve demanding problems. We strive to develop novel architectures, applications and design techniques for problems combining signal processing and machine learning.

## 23rd Oct 2019, Lunchtime Tutorial

Python Map-Reduce tutorial, Computer Engineering Lab

This Python tutorial introduces the map-reduce programming model. Map-reduce can be used to simplify code as well as to parallelise computer programs. It is commonly used not only in Python programs, but also in reconfigurable computing and large-scale computing.

## 30th Sep 2019, Lecture

Multipliers for FPGA Machine Learning Applications, Fudan University

FPGA implementations of machine learning algorithms have received considerable interest, and multipliers are a key component for their efficient implementation. In this talk, four examples of our recent research efforts in customisation of FPGA architecture and designs using different approaches to multiplication be described. As the multipliers become more specialised, higher performance gains are possible.

First we describe a two-speed multiplier (TSM) which offers improved average clock rate over a conventional multiplier by carefully considering the delay of two critical paths in a serial-parallel modified radix-4 Booth multiplier. Secondly, our joint research with Fudan University on a flexible precision, run-time decomposable DSP architecture for CNN implementations which optimises Precision, Interconnect, and Reuse (PIR-DSP) is presented. The resulting design offers a 6× improvement in multiply-accumulate operations per DSP in the 9 × 9-bit case, 12× for 4 × 4 bits, and 24× for 2 × 2 bits. Thirdly, we present AddNet which shows that reconfigurable constant coefficient multipliers (RCCMs) offer an alternative to low precision arithmetic for saving area. Finally, we describe a ternary neural network generator which can generate fully unrolled designs. The implementation of a VGG-style network achieves 90.9% accuracy and 122 k frames per second, with a latency of only 29 μs, which is the fastest CNN inference implementation reported so far on an FPGA.

## 10th May 2019, Lecture

Efficient FPGA implementations of Machine Learning Algorithms, San Jose State University

FPGA implementations of machine learning algorithms have been shown to be extremely efficient when the problem fits entirely on the FPGA but it remains a challenge to scale to problems of interest to industry. In this talk, our recent research on how to increase the capacity of existing approaches will be described.

In the first part of the talk we will describe an implementation of the Fastfood algorithm for scaling up online kernel methods. By utilizing the theory of random projections, problems with 1000x larger input dimension and dictionary size can be solved. A systolic implementation that operates at 500 MHz clock frequency was achieved. The next part of the talk describes how the resource requirements in convolutional neural networks can be reduced using symmetric quantisation (SYQ), which achieves state of the art accuracy for binary and ternary weights. Finally, our work on new DSP-block architectures within FPGAs, optimised for energy-efficient embedded machine learning will be presented.

## 7th December 2018, Lecture

Large Scale FPGA Implementations of Machine Learning Algorithms, Tokyo Institute of Technology

*FPGA implementations of machine learning algorithms have been shown to be extremely efficient when the problem fits entirely on the FPGA but it remains a challenge to scale to problems of interest to industry. In this talk, our recent research on how to increase the capacity of existing approaches will be described.*

*In the first part of the talk we will describe an implementation of the Fastfood algorithm for scaling up online kernel methods. By utilizing the theory of random projections, problems with 1000x larger input dimension and dictionary size can be solved. A systolic implementation that operates at 500 MHz clock frequency was achieved. The next part of the talk describes how the resource requirements in convolutional neural networks can be reduced using symmetric quantisation (SYQ), which achieves state of the art accuracy for binary and ternary weights. Finally, our work on an open-source matrix multiplication library on the Xeon+FPGA platform is described. This allows a number of different precisions down to binary, and additional support for deep learning applications is provided. Performance approaching the maximum of 14 Xeon cores and an Arria 10 FPGA was achieved.*

## 26 November 2018, Lecture

Fixed-point FPGA implementations of Machine Learning Algorithms, Department of Computing, Imperial College

*Abstract: FPGA implementations of machine learning algorithms have been shown to be extremely efficient, particularly when fixed-point arithmetic can be utilised to simplify the datapath of a hardware implementation. In this talk, our recent research on how accuracy and performance of fixed-point implementations of neural networks can be improved will be presented.*

*In the first part of the talk we describe how the resource requirements in convolutional neural networks can be reduced using symmetric quantisation (SYQ), which achieves state of the art accuracy for binary and ternary weights. Next, a variant of the serial-parallel modified radix-4 Booth multiplier that adds only the non-zero Booth encodings and skips over the zero operations, achieving a 1.4x Area-Time improvement over the standard parallel Booth multiplier is presented. Finally, we describe a system implementation of a long short-term memory (LSTM) based spectral prediction accelerator for physical layer radio frequency applications, with latency below 5 microseconds.*

## 29 October 2018, Lecture

Long short-term memory for radio frequency spectral prediction and its real-time FPGA implementation, MILCOM Los Angeles

## 26 September 2018, Lecture

Teaching and Research at the University of Sydney, Harbin Institute of Technology

## 26 September 2018, Lecture

Reconfigurable Computing, Harbin Institute of Technology Short Graduate Course

## 25 September 2018, Lecture

Discover your future in engineering and technology, D & LT Consultants, Harbin

## 24 July 2018, Lecture

Quantisation, Performance-Aware Programming with Application Accelerators (PAPAA), Hong Kong

## 30 November 2017, Lecture

FPGA-based Implementations of Machine Learning Algorithms and the EPIC Approach, Imperial College, London.

Abstract - Recent progress in machine learning (ML) has led to an ability to solve difficult problems with unprecedented accuracy. Unfortunately, implementations using conventional software-based technologies such as central processing units, digital signal processors and graphics processing units suffer from high latency and high power consumption. In this talk, it will be explained how we have been applying Exploration, Parallelism, Integration and Customisation (EPIC) design techniques to FPGA implementations of online kernel methods, low-precision convolutional neural networks, and real-time processing of radio-frequency signals._

## 21 October 2017, Keynote Speech

FPGA-based Machine Learning for Electronic Measurement and Instruments, IEEE International Conference on Electronic Measurement and Instruments (ICEMI) Conference, Yangzhou, China

## 24 October 2016, Lecture

Architectures for the FPGA Implementation of Online Kernel Methods (link to video), Microsoft Research Talks, Redmond

In machine learning, traditional linear prediction techniques are well understood and methods for their efficient solution have been developed. Many real-world applications are better modelled using non-linear techniques, which often have high computational requirements. Kernel methods utilise linear methods in a non-linear feature space and combine the advantages of both. Commonly used kernel methods include the support vector machine (SVM), Gaussian processes and regularisation networks. These are batch-based, and a global optimisation is conducted over all input exemplars to create a model. In contrast, online methods, such as the kernel recursive least squares (KRLS) algorithm, update the state in a recursive and incremental fashion upon receiving a new exemplar. Although not as extensively studied as batch methods, online approaches are advantageous when throughput and latency are critical. In this talk I will describe efforts in the Computer Engineering Laboratory to produce high-performance FPGA-based implementations of online kernel methods. These have included: (1) a microcoded vector processor optimised for kernel methods; (2) a fully pipelined implementation of kernel normalised least mean squares which achieves 160 GFLOPS; (3) an implementation of Naive Online regularised Risk Minimization Algorithm (NORMA) which uses “braiding” to resolve data hazards and reduce latency by an order of magnitude; and (4) a distributed kernel recursive least squares algorithm which constructs a compact model while enabling massive parallelism.

## 25 May 2015, Lecture

Integration of Nanoscale Structures with Electronics, Australian Institute for Nanoscale Science and Technology Accelerator Seminar, The University of Sydney

## 27 November 2015, Lecture

Architectures for the FPGA Implementation of Online Kernel Methods, Xilinx Research Labs, Dublin, Ireland

## 1 July 2014, Lecture

MCALIB - Measuring Sensitivity to Rounding Error using Monte Carlo Programming, Microsoft Research Asia, Beijing, China

## 25 October 2013, Lecture

Computer Engineering Lab, National ICT Australia (NICTA), Sydney

## 14 December 2007, Lecture

10 Sonnets Concerning Field Programmable Technology, International Conference on Field Programmable Technology Special Session on What is Exciting about Field Programmable Technology