Themis Prodromakis

Cristian Sestito, Shady Agwa, Themis Prodromakis

Modern hardware architectures for Convolutional Neural Networks (CNNs), other than targeting high performance, aim at dissipating limited energy. Reducing the data movement cost between the computing cores and the memory is a way to mitigate the energy consumption. Systolic arrays are suitable architectures to achieve this objective: they use multiple processing elements that communicate each other to maximize data utilization, based on proper dataflows like the weight stationary and row stationary. Motivated by this, we have proposed TrIM, an innovative dataflow based on a triangular movement of inputs, and capable to reduce the number of memory accesses by one order of magnitude when compared to state-of-the-art systolic arrays. In this paper, we present a TrIM-based hardware architecture for CNNs. As a showcase, the accelerator is implemented onto a Field Programmable Gate Array (FPGA) to execute the VGG-16 and AlexNet CNNs. The architecture achieves a peak throughput of 453.6 Giga Operations per Second, outperforming a state-of-the-art row stationary systolic array up to ~3x in terms of memory accesses, and being up to ~11.9x more energy-efficient than other FPGA accelerators.

Cristian Sestito, Shady Agwa, Themis Prodromakis

In order to follow the ever-growing computational complexity and data intensity of state-of-the-art AI models, new computing paradigms are being proposed. These paradigms aim at achieving high energy efficiency by mitigating the Von Neumann bottleneck that relates to the energy cost of moving data between the processing cores and the memory. Convolutional Neural Networks (CNNs) are susceptible to this bottleneck, given the massive data they have to manage. Systolic arrays (SAs) are promising architectures to mitigate data transmission cost, thanks to high data utilization of Processing Elements (PEs). These PEs continuously exchange and process data locally based on specific dataflows (such as weight stationary and row stationary), in turn reducing the number of memory accesses to the main memory. In SAs, convolutions are managed either as matrix multiplications or exploiting the raster-order scan of sliding windows. However, data redundancy is a primary concern affecting area, power, and energy. In this paper, we propose TrIM: a novel dataflow for SAs based on a Triangular Input Movement and compatible with CNN computing. TrIM maximizes the local input utilization, minimizes the weight data movement, and solves the data redundancy problem. Furthermore, TrIM does not incur the significant on-chip memory penalty introduced by the row stationary dataflow. When compared to state-of-the-art SA dataflows, the high data utilization offered by TrIM guarantees ~10X less memory access. Furthermore, considering that PEs continuously overlap multiplications and accumulations, TrIM achieves high throughput (up to 81.8% higher than row stationary), other than requiring a limited number of registers (up to 15.6X fewer registers than row stationary).

Kieran Woodward, Eiman Kanjo, Georgios Papandroulidakis et al.

In recent years, the development of smart edge computing systems to process information locally is on the rise. Many near-sensor machine learning (ML) approaches have been implemented to introduce accurate and energy efficient template matching operations in resource-constrained edge sensing systems, such as wearables. To introduce novel solutions that can be viable for extreme edge cases, hybrid solutions combining conventional and emerging technologies have started to be proposed. Deep Neural Networks (DNN) optimised for edge application alongside new approaches of computing (both device and architecture -wise) could be a strong candidate in implementing edge ML solutions that aim at competitive accuracy classification while using a fraction of the power of conventional ML solutions. In this work, we are proposing a hybrid software-hardware edge classifier aimed at the extreme edge near-sensor systems. The classifier consists of two parts: (i) an optimised digital tinyML network, working as a front-end feature extractor, and (ii) a back-end RRAM-CMOS analogue content addressable memory (ACAM), working as a final stage template matching system. The combined hybrid system exhibits a competitive trade-off in accuracy versus energy metric with $E_{front-end}$ = $96.23 nJ$ and $E_{back-end}$ = $1.45 nJ$ for each classification operation compared with 78.06$μ$J for the original teacher model, representing a 792-fold reduction, making it a viable solution for extreme edge applications.

Sachin Maheshwari, Alexander Serb, Christos Papavassiliou et al.

In the quest for low power, bio-inspired computation both memristive and memcapacitive-based Artificial Neural Networks (ANN) have been the subjects of increasing focus for hardware implementation of neuromorphic computing. One step further, regenerative capacitive neural networks, which call for the use of adiabatic computing, offer a tantalising route towards even lower energy consumption, especially when combined with `memimpedace' elements. Here, we present an artificial neuron featuring adiabatic synapse capacitors to produce membrane potentials for the somas of neurons; the latter implemented via dynamic latched comparators augmented with Resistive Random-Access Memory (RRAM) devices. Our initial 4-bit adiabatic capacitive neuron proof-of-concept example shows 90% synaptic energy saving. At 4 synapses/soma we already witness an overall 35% energy reduction. Furthermore, the impact of process and temperature on the 4-bit adiabatic synapse shows a maximum energy variation of 30% at 100 degree Celsius across the corners without any functionality loss. Finally, the efficacy of our adiabatic approach to ANN is tested for 512 & 1024 synapse/neuron for worst and best case synapse loading conditions and variable equalising capacitance's quantifying the expected trade-off between equalisation capacitance and range of optimal power-clock frequencies vs. loading (i.e. the percentage of active synapses).

Alexander Serb, Themistoklis Prodromakis

The field of artificial intelligence (AI) represents an enormous endeavour of humankind that is currently transforming our societies down to their very foundations. Its task, building truly intelligent systems, is underpinned by a vast array of subfields ranging from the development of new electronic components to mathematical formulations of highly abstract and complex reasoning. This breadth of subfields renders it often difficult to understand how they all fit together into a bigger picture and hides the multi-faceted, multi-layered conceptual structure that in a sense can be said to be what AI truly is. In this perspective we propose a system of five levels/layers of abstraction that underpin many AI implementations. We further posit that each layer is subject to a complexity-performance trade-off whilst different layers are interlocked with one another in a control-complexity trade-off. This overview provides a conceptual map that can help to identify how and where innovation should be targeted in order to achieve different levels of functionality, assure them for safety, optimise performance under various operating constraints and map the opportunity space for social and economic exploitation.

A. Serb, I. Kobyzev, J. Wang et al.

One of the main, long-term objectives of artificial intelligence is the creation of thinking machines. To that end, substantial effort has been placed into designing cognitive systems; i.e. systems that can manipulate semantic-level information. A substantial part of that effort is oriented towards designing the mathematical machinery underlying cognition in a way that is very efficiently implementable in hardware. In this work we propose a 'semi-holographic' representation system that can be implemented in hardware using only multiplexing and addition operations, thus avoiding the need for expensive multiplication. The resulting architecture can be readily constructed by recycling standard microprocessor elements and is capable of performing two key mathematical operations frequently used in cognition, superposition and binding, within a budget of below 6 pJ for 64- bit operands. Our proposed 'cognitive processing unit' (CoPU) is intended as just one (albeit crucial) part of much larger cognitive systems where artificial neural networks of all kinds and associative memories work in concord to give rise to intelligence.

Chuan Kai Kenneth. Lim, T. Prodromakis

Computing the relative motion of objects is an important navigation task that we routinely perform by relying on inherently unreliable biological cells in the retina. The non-linear and adaptive response of memristive devices make them excellent building blocks for realizing complex synaptic-like architectures that are common in the human retina. Here, we introduce a novel memristive thresholding scheme that facilitates the detection of moving edges. In addition, a double-layered 3-D memristive network is employed for modeling the motion computations that take place in both the Outer Plexiform Layer (OPL) and Inner Plexiform Layer (IPL) that enables the detection of on-center and off-center transient responses. Applying the transient detection results, it is shown that it is possible to generate an estimation of the speed and direction a moving object.