Topkima-Former: Low-energy, Low-Latency Inference for Transformers using top-k In-memory ADC

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Dong, Shuai, Yang, Junyi, Peng, Xiaoqi, Shang, Hongyang, Ke, Ye, Yang, Xiaofeng, Liu, Hongjie, Basu, Arindam
Format: Preprint
Veröffentlicht: 2024
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866916489109438464
author Dong, Shuai
Yang, Junyi
Peng, Xiaoqi
Shang, Hongyang
Ke, Ye
Yang, Xiaofeng
Liu, Hongjie
Basu, Arindam
author_facet Dong, Shuai
Yang, Junyi
Peng, Xiaoqi
Shang, Hongyang
Ke, Ye
Yang, Xiaofeng
Liu, Hongjie
Basu, Arindam
contents Transformer model has gained prominence as a popular deep neural network architecture for neural language processing (NLP) and computer vision (CV) applications. However, the extensive use of nonlinear operations, like softmax, poses a performance bottleneck during transformer inference and comprises up to 40% of the total latency. Hence, we propose innovations at the circuit, architecture, and algorithm levels to accelerate the transformer. At the circuit level, we propose topkima-combining top-k activation selection with in-memory ADC (IMA) to implement a low-energy and low-latency softmax without any sorting latency. Only the k largest activations are sent to the softmax calculation block, reducing the huge computational cost of softmax. Using a modified training scheme with top-k only in the forward pass, experimental results demonstrate only a 0.4% to 1.2% reduction in accuracy across ViT, distilBERT, and BERT-base models when evaluated on CIFAR-10, CIFAR-100, and SQuAD datasets with k=5. At the architecture level, an improved scale-free technique is introduced to reduce the computational cost of attention. The combined system, dubbed Topkima-Former, enhances 1.8x-84x speedup and 1.3x-35x energy efficiency (EE) over prior In-memory computing (IMC) accelerators. Compared to a conventional softmax macro and a digital top-k (Dtopk) softmax macro, our proposed tokima softmax macro achieves about 15x and 8x faster speed respectively.
format Preprint
id arxiv_https___arxiv_org_abs_2411_13050
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Topkima-Former: Low-energy, Low-Latency Inference for Transformers using top-k In-memory ADC
Dong, Shuai
Yang, Junyi
Peng, Xiaoqi
Shang, Hongyang
Ke, Ye
Yang, Xiaofeng
Liu, Hongjie
Basu, Arindam
Hardware Architecture
Transformer model has gained prominence as a popular deep neural network architecture for neural language processing (NLP) and computer vision (CV) applications. However, the extensive use of nonlinear operations, like softmax, poses a performance bottleneck during transformer inference and comprises up to 40% of the total latency. Hence, we propose innovations at the circuit, architecture, and algorithm levels to accelerate the transformer. At the circuit level, we propose topkima-combining top-k activation selection with in-memory ADC (IMA) to implement a low-energy and low-latency softmax without any sorting latency. Only the k largest activations are sent to the softmax calculation block, reducing the huge computational cost of softmax. Using a modified training scheme with top-k only in the forward pass, experimental results demonstrate only a 0.4% to 1.2% reduction in accuracy across ViT, distilBERT, and BERT-base models when evaluated on CIFAR-10, CIFAR-100, and SQuAD datasets with k=5. At the architecture level, an improved scale-free technique is introduced to reduce the computational cost of attention. The combined system, dubbed Topkima-Former, enhances 1.8x-84x speedup and 1.3x-35x energy efficiency (EE) over prior In-memory computing (IMC) accelerators. Compared to a conventional softmax macro and a digital top-k (Dtopk) softmax macro, our proposed tokima softmax macro achieves about 15x and 8x faster speed respectively.
title Topkima-Former: Low-energy, Low-Latency Inference for Transformers using top-k In-memory ADC
topic Hardware Architecture
url https://arxiv.org/abs/2411.13050