Survey of Machine Learning Accelerators

Reuther, Albert; Michaleas, Peter; Jones, Michael; Gadepally, Vijay; Samsi, Siddharth; Kepner, Jeremy

doi:10.1109/hpec43674.2020.9286149

Cited by 111 publications

(57 citation statements)

References 38 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…All weights implemented in our system are passive, making the system continuously operate at the DC power-levels of its components. > 10 6 inferences per second with 100 TBit/s input data rate at 1…10 W power consumption are therefore realistic, which is orders of magnitude beyond any electronic special purpose ANN computing hardware [40].…”

Section: Discussionmentioning

confidence: 99%

A complete, parallel, and autonomous photonic neural network in a semiconductor multimode laser

Porté

Skalli

Haghighi

et al. 2021

Emerging Topics in Artificial Intelligence (ETAI) 2021

View full text Add to dashboard Cite

Neural networks are one of the disruptive computing concepts of our time. However, they fundamentally differ from classical, algorithmic computing. These differences result in equally fundamental, severe and relevant challenges for neural network computing using current computing substrates. Neural networks urge for parallelism across the entire processor and for a co-location of memory and arithmetic, i.e. beyond von Neumann architectures. Parallelism in particular made photonics a highly promising platform, yet until now scalable and integratable concepts are scarce. Here, we demonstrate for the first time how a fully parallel and fully implemented photonic neural network can be realized by spatially multiplexing neurons across the complex optical near-field of a semiconductor multimode laser. Discrete spatial sampling defines ∼90 nodes on the surface of a large-area vertical cavity surface emitting laser that is optically injected with the artificial neural networks input information. Importantly, all neural network connections are realized in hardware, and our processor produces results without pre-or post-processing. Input and output weights are realized via the complex transmission matrix of a multimode fiber and a digital micro-mirror array, respectively. We train the readout weights to perform 2-bit header recognition, a 2-bit XOR logical function and 2-bit digital to analog conversion, and obtain < 0.9 × 10 −3 and 2.9 × 10 −2 error rates for digit recognition and XOR, respectively. Finally, the digital to analog conversion can be realized with a standard deviation of only 5.4 × 10 −2 . Crucially, our proof-of-concept system is scalable to much larger sizes and to bandwidths in excess of 20 GHz.

show abstract

Section: Discussionmentioning

confidence: 99%

A complete, parallel, and autonomous photonic neural network in a semiconductor multimode laser

Porté

Skalli

Haghighi

et al. 2021

Emerging Topics in Artificial Intelligence (ETAI) 2021

View full text Add to dashboard Cite

show abstract

“…Because the basic algorithm does not change much over different applications, CNNs in essence provide a universal solution to many compute tasks. This makes them a highly eligible target for dedicated hardware solutions, and as such has inspired the design of many CNN hardware accelerators [2].…”

Section: Related Workmentioning

confidence: 99%

“…In the early years this kept the execution of CNNs confined to data centres, as evaluation on available general purpose, embedded processors required too much energy to be practical in mobile, energy constrained devices. To overcome this, many dedicated CNN accelerators have been proposed since to bring CNNs to the edge, and in general reduce CNN energy consumption [2]. In modern technology nodes the main challenge in achieving a high energy efficiency for such accelerators is not so much the compute complexity, but rather the required memory accesses.…”

mentioning

confidence: 99%

ConvFusion: A Model for Layer Fusion in Convolutional Neural Networks

Waeijen¹,

Sioutas²,

Peemen

et al. 2021

IEEE Access

View full text Add to dashboard Cite

The superior accuracy and appealing universality of convolutional neural networks (CNNs) as a generic algorithm for many classification tasks have made the design of energy efficient CNN accelerators an important topic in both academia and industry. Of particular interest in the design and use of CNN accelerators is the scheduling of the computational workload, which can have a major impact on the quality of the final design. The many inherently independent operations in CNNs result in a vast scheduling space however, rendering the selection of the optimal schedule(s) non-trivial. To aid in this complex task, this work introduces a generic mathematical cost model of the external memory accesses, internal memory footprint, and compute load for CNN execution schedules. The model enables fast exploration of the scheduling space, including loop tiling, loop reordering, explicit data transfer scheduling, recomputation, and, crucially, layer fusion, which recently has attracted interest as a method to reduce external memory accesses. An accompanying open source tool is released to perform schedule space exploration for CNNs using the introduced cost model. Leveraging the code generation capabilities of this tool the proposed model is validated on six real world networks, demonstrating that layer fusion can reduce the external memory accesses by more than two orders of magnitude compared to the best non-fused schedules. Confusing at first glance however, a high-level energy analysis shows that the practical benefits of layer fusion may be overestimated if other parts of the system are not tuned accordingly.

show abstract

“…The higher feedback speed will result in higher yield production. Here, we can accelerate the abnormal detection because we can apply high performance processors on the server side with Graphics Processing Units (GPUs) [ 13 ] and Artificial Intelligence (AI) processors [ 14 ]. On the other hand, the communication part of the system includes the lossy compression and the transfer via a network media.…”

Section: Introductionmentioning

confidence: 99%

Stream-Based Visually Lossless Data Compression Applying Variable Bit-Length ADPCM Encoding

Yamagiwa

Ichinomiya

2021

Sensors

View full text Add to dashboard Cite

Video applications have become one of the major services in the engineering field, which are implemented by server–client systems connected via the Internet, broadcasting services for mobile devices such as smartphones and surveillance cameras for security. Recently, the majority of video encoding mechanisms to reduce the data rate are mainly lossy compression methods such as the MPEG format. However, when we consider special needs for high-speed communication such as display applications and object detection ones with high accuracy from the video stream, we need to address the encoding mechanism without any loss of pixel information, called visually lossless compression. This paper focuses on the Adaptive Differential Pulse Code Modulation (ADPCM) that encodes a data stream into a constant bit length per data element. However, the conventional ADPCM does not have any mechanism to control dynamically the encoding bit length. We propose a novel ADPCM that provides a mechanism with a variable bit-length control, called ADPCM-VBL, for the encoding/decoding mechanism. Furthermore, since we expect that the encoded data from ADPCM maintains low entropy, we expect to reduce the amount of data by applying a lossless data compression. Applying ADPCM-VBL and a lossless data compression, this paper proposes a video transfer system that controls throughput autonomously in the communication data path. Through evaluations focusing on the aspects of the encoding performance and the image quality, we confirm that the proposed mechanisms effectively work on the applications that needs visually lossless compression by encoding video stream in low latency.

show abstract

Survey of Machine Learning Accelerators

Cited by 111 publications

References 38 publications

A complete, parallel, and autonomous photonic neural network in a semiconductor multimode laser

A complete, parallel, and autonomous photonic neural network in a semiconductor multimode laser

ConvFusion: A Model for Layer Fusion in Convolutional Neural Networks

Stream-Based Visually Lossless Data Compression Applying Variable Bit-Length ADPCM Encoding

Contact Info

Product

Resources

About