ECIM: Exponent Computing in Memory for an Energy-Efficient Heterogeneous Floating-Point DNN Training Processor

Lee, Juhyoung; Kim, Jihoon; Jo, Wooyoung; Kim, Sangjin; Yoo, Hoi‐Jun

doi:10.1109/mm.2021.3096236

Cited by 10 publications

(4 citation statements)

References 6 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Yet most existing SRAM-based IMC structures have low data density 8 . The data density in [41,42] is lower than 0.3Mb mm −2 . This inevitably causes more off-chip memory communication during training, and can hardly scale out for large-scale training tasks at low cost.…”

Section: Introductionmentioning

confidence: 79%

“…Recently, there are several works employing floating-point formats for IMC designs: Tu et al [41] developed a digital processor handling GeMM between integers and FP numbers, reaching 14 TFLOPS/W efficiency computing GeMM with BFloat16 format. Lee et al [42] designed an SRAM-based IMC circuit processing GeMM between BFloat16, which separates exponent and fraction storage with the best computing efficiency at 1.43 TFLOPS W −1 ; Lee et al [43] give a DRAM-based near-memory computing design with high throughput at 1 TFLOPS per chip. Nevertheless, efficient in-memory simultaneous processing of mantissa and exponents are still not explored.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Hadamard product-based in-memory computing design for floating point neural network training

Fan

Tao

et al. 2023

Neuromorph. Comput. Eng.

View full text Add to dashboard Cite

Deep neural networks (DNNs) are one of the key fields of machine learning. It requires considerable computational resources for cognitive tasks. As a novel technology to perform computing inside/near memory units, in-memory computing (IMC) significantly improves computing efficiency by reducing the need for repetitive data transfer between the processing and memory units. However, prior IMC designs mainly focus on the acceleration for DNN inference. DNN training with the IMC hardware has rarely been proposed. The challenges lie in the requirement of DNN training for high precision (e.g. floating point) and various operations of tensors (e.g. inner and outer products). These challenges call for the IMC design with new features. This paper proposes a novel Hadamard product-based IMC design for floating point DNN training. Our design consists of multiple compartments, which are the basic units for the matrix element-wise processing. We also develop BFloat16 post-processing circuits and fused adder trees, laying the foundation for IMC floating point processing. Based on the proposed circuit scheme, we reformulate the back-propagation training algorithm for the convenience and efficiency of the IMC execution. The proposed design is implemented with commercial 28nm technology process design kits and benchmarked with widely used neural networks. We model the influence of the circuit structural design parameters and provide an analysis framework for design space exploration. Our simulation validates that MobileNet training with the proposed IMC scheme saves 91.2% in energy and 13.9% in time versus the same task with NVIDIA GTX 3060 GPU. The proposed IMC design has a data density of 769.2Kb/mm2 with the floating point processing circuits included, showing a 3.5× improvement than the prior floating point IMC designs.

show abstract

Section: Introductionmentioning

confidence: 79%

Section: Introductionmentioning

confidence: 99%

Hadamard product-based in-memory computing design for floating point neural network training

Fan

Tao

et al. 2023

Neuromorph. Comput. Eng.

View full text Add to dashboard Cite

show abstract

“…The architecture uses bfloat16DNN to train the processor to optimize the exponential computation and to improve energy efficiency and reduce memory consumption. The resulting computational architecture achieves an energy efficiency of 13.7 TFLOPS/W when processing data [6]. Wang C et al developed a DNN-based optimized read voltage value strategy for expressing the relationship between voltage distribution and read voltage threshold.…”

Section: Related Workmentioning

confidence: 99%

“…In equation (3.12), p is the quantity of neurons in the hidden layer, n is the quantity of neurons in the input layer, and a is the mediation parameter, which takes values in the range of [1][2][3][4][5][6][7][8][9][10]. To sum up, this research first carries out the assessment index confirmation of the TSC risk of enterprises, then data and normalization of the index, then constructs a multi-level DNN model to calculate the characteristics of the processed index, and finally judges the TSC risk degree of the enterprise according to the calculation results.…”

Section: Construction Of a Multi-level Dnn-based Risk Prediction Mode...mentioning

confidence: 99%

A Multi-Level Deep Neural Network-Based Tourism Supply Chain Risk Management Study

2024

SCPE

View full text Add to dashboard Cite

With the rapid advancement of the tourism, the capital demand of tourism enterprises has gradually risen, but the confusion of market management has increased the difficulty of risk assessment of tourism enterprises by financial institutions, which has led to difficulties in financing tourism enterprises and seriously hindered their development. On this basis, this research first analyses the risk structure of the tourism supply chain (TSC) from the financial aspect and establishes a relevant risk assessment system. After confirming the assessment indexes, the data is dativized and normalized, and finally a multi-level deep neural network (DNN) is used to construct a TSC risk prediction model to calculate the transformed indexes and assess the risk degree of the enterprise according to the results. The experimental results indicate that the model has the best performance when the H-Net hidden layer is 3 layers and the L-Net hidden layer is 4 layers, and its accuracy reaches 93.35%, sensitivity reaches 84.13%, convergence starts at 25 iterations, the final loss value is only 0.8, and the predicted and the real value error is within 2.5%. Therefore, the multi-level DNN model constructed in this experiment has certain application value in TSC risk management.

show abstract

PIM for ML Training

Heo

Kim

2012

Processing-in-Memory for AI

View full text Add to dashboard Cite

ECIM: Exponent Computing in Memory for an Energy-Efficient Heterogeneous Floating-Point DNN Training Processor

Cited by 10 publications

References 6 publications

Hadamard product-based in-memory computing design for floating point neural network training

Hadamard product-based in-memory computing design for floating point neural network training

A Multi-Level Deep Neural Network-Based Tourism Supply Chain Risk Management Study

PIM for ML Training

Contact Info

Product

Resources

About