AI and ML Accelerator Survey and Trends

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

doi:10.1109/hpec55821.2022.9926331

Cited by 36 publications

(16 citation statements)

References 47 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…However, despite all the advances, empirical evidence based on commercial AI accelerators indicate plateauing of energy efficiency at 10 TOPS/W or 100 fJ/Operation. 221 This is attributed to the energy costs associated with data movement, as indicated in Figure 18a. MVM operations cost tens of fJ whereas accessing on-chip SRAM costs about 1pJ/byte.…”

Section: Development Of Computing-in-memory Based Ai Acceleratorsmentioning

confidence: 99%

“…Digital accelerators typically contain arrays of processor elements (PEs) that can perform several MVM operations in a pipeline-parallel fashion. However, despite all the advances, empirical evidence based on commercial AI accelerators indicate plateauing of energy efficiency at 10 TOPS/W or 100 fJ/Operation . This is attributed to the energy costs associated with data movement, as indicated in Figure a.…”

Section: Computing-in-memory System For Ai Acceleratorsmentioning

confidence: 99%

See 1 more Smart Citation

Recent Advances and Future Prospects for Memristive Materials, Devices, and Systems

et al. 2023

View full text Add to dashboard Cite

Memristive technology has been rapidly emerging as a potential alternative to traditional CMOS technology, which is facing fundamental limitations in its development. Since oxide-based resistive switches were demonstrated as memristors in 2008, memristive devices have garnered significant attention due to their biomimetic memory properties, which promise to significantly improve power consumption in computing applications. Here, we provide a comprehensive overview of recent advances in memristive technology, including memristive devices, theory, algorithms, architectures, and systems. In addition, we discuss research directions for various applications of memristive technology including hardware accelerators for artificial intelligence, in-sensor computing, and probabilistic computing. Finally, we provide a forward-looking perspective on the future of memristive technology, outlining the challenges and opportunities for further research and innovation in this field. By providing an up-to-date overview of the state-of-the-art in memristive technology, this review aims to inform and inspire further research in this field.

show abstract

Section: Development Of Computing-in-memory Based Ai Acceleratorsmentioning

confidence: 99%

Section: Computing-in-memory System For Ai Acceleratorsmentioning

confidence: 99%

Recent Advances and Future Prospects for Memristive Materials, Devices, and Systems

et al. 2023

View full text Add to dashboard Cite

show abstract

“…Litz wire has also been found particularly suitable for constructing ULF rf receiver coil (68)(69)(70). On the other hand, new computational strategies will be continuously developed to advance ULF image quality (e.g., higher spatial resolution by higher SR factor) by exploiting rapidly evolving DL algorithms and architectures as well as ever increasing computing power and large-scale human MRI data availability (59,71,72). Future efforts shall also encompass the experimental assessment and optimization of ULF data acquisition and DL image reconstruction to yield optimal trade-offs between image fidelity, resolution, contrast, scan time, and cost for each specific applications.…”

Section: Challengesmentioning

confidence: 99%

Deep learning enabled fast 3D brain MRI at 0.055 tesla

Man,

Lau,

et al. 2023

Sci. Adv.

View full text Add to dashboard Cite

In recent years, there has been an intensive development of portable ultralow-field magnetic resonance imaging (MRI) for low-cost, shielding-free, and point-of-care applications. However, its quality is poor and scan time is long. We propose a fast acquisition and deep learning reconstruction framework to accelerate brain MRI at 0.055 tesla. The acquisition consists of a single average three-dimensional (3D) encoding with 2D partial Fourier sampling, reducing the scan time of T1- and T2-weighted imaging protocols to 2.5 and 3.2 minutes, respectively. The 3D deep learning leverages the homogeneous brain anatomy available in high-field human brain data to enhance image quality, reduce artifacts and noise, and improve spatial resolution to synthetic 1.5-mm isotropic resolution. Our method successfully overcomes low-signal barrier, reconstructing fine anatomical structures that are reproducible within subjects and consistent across two protocols. It enables fast and quality whole-brain MRI at 0.055 tesla, with potential for widespread biomedical applications.

show abstract

“…Introduction of interaction between weights is not strictly speaking necessary to satisfy the condition (5). Since magnetic moments directions are subjected to normalization condition S 2 j = 1, the energy functional is always bounded, hence it has at least one minimum.…”

Section: Spin Boltzmann Machinementioning

confidence: 99%

“…At the moment computers can solve many artificial intelligence (AI) tasks, however the energy efficiency of the organic brain is still much higher. The need to reduce energy consumption leads to the emergence of dedicated AI accelerators [5]. The widespread digital computers are good for precise computations, but analog devices and stochastic computing better suit the needs of ML [6].…”

Section: Introductionmentioning

confidence: 99%

Spin Boltzmann machine

Lobanov¹

2022

Nanosystems: Phys. Chem. Math.

View full text Add to dashboard Cite

Boltzmann machine (BM) is a recurrent network, which has a wide range of applications in machine learning (ML) including dimensionality reduction, feature learning and classification. Standard BM is described by the Ising model and can be implemented as a spin ice based device. Such hardware implementation is faster and more energy efficient than a simulation on digital computers. At the moment, a hardware BM is a single purpose device designed on digital computers for a specific task. In the paper we propose a generalized BM capable of fitting parameters by demonstration of training examples, which is done completely inside the spintronic device. Our generalization is based on the Heisenberg model, which is more accurate than the Ising model for spin ice. We show that for some systems minimization of Kullback-Leibler divergence during training of BM is equivalent to minimization of free energy with respect to the biases of the units, hence training of the ML model can be done by energy dissipation. We include the biases as degrees of freedom of the device, whose dynamics is described by the same Landau-Lifschitz-Gilbert equation as for spins representing units of BM. The demonstration of samples from the training set is done by fixing inputs and outputs according to ground truth. The training samples are remembered by the machine becoming minima on the energy landscape implementing a kind of long-term potentiation. The performance of the proposed machine is compared with a single layer perceptron artificial neural network and with a Bernoulli restricted BM on a binary classification problem. KEYWORDS spectral gap, quantum graph, Schr ödinger operator, discrete spectrum.

show abstract

AI and ML Accelerator Survey and Trends

Cited by 36 publications

References 47 publications

Recent Advances and Future Prospects for Memristive Materials, Devices, and Systems

Recent Advances and Future Prospects for Memristive Materials, Devices, and Systems

Deep learning enabled fast 3D brain MRI at 0.055 tesla

Spin Boltzmann machine

Contact Info

Product

Resources

About