Transformer models have emerged as the state-of-the-art in many natural language processing and computer vision applications due to their capability of attending to longer sequences of tokens and supporting parallel processing more efficiently. Nevertheless, the training and inference of transformer models are computationally expensive and memory intensive. Meanwhile, utilizing the sparsity in deep learning models has proven to be an effective approach to alleviate the computation challenge as well as help to fit large models in edge devices. As high-performance CPUs and GPUs are generally not flexible enough to explore low-level sparsity, a number of specialized hardware accelerators have been proposed for transformer models. This paper provides a comprehensive review of hardware transformer accelerators that have been proposed to explore sparsity for computation and memory optimizations. We classify existing works based on the strategies of utilizing sparsity and identify their pros and cons in those strategies. Based on our analysis, we point out promising directions and recommendations for future works on improving the effective sparse execution of transformer hardware accelerators.
more »
« less
22.9 A 12nm 18.1TFLOPs/W Sparse Transformer Processor with Entropy-Based Early Exit, Mixed-Precision Predication and Fine-Grained Power Management
Large language models have substantially advanced nuance and context understanding in natural language processing (NLP), further fueling the growth of intelligent conversational interfaces and virtual assistants. However, their hefty computational and memory demands make them potentially expensive to deploy on cloudless edge platforms with strict latency and energy requirements. For example, an inference pass using the state-of-the-art BERT-base model must serially traverse through 12 computationally intensive transformer layers, each layer containing 12 parallel attention heads whose outputs concatenate to drive a large feed-forward network. To reduce computation latency, several algorithmic optimizations have been proposed, e.g., a recent algorithm dynamically matches linguistic complexity with model sizes via entropy-based early exit. Deploying such transformer models on edge platforms requires careful co-design and optimizations from algorithms to circuits, where energy consumption is a key design consideration.
more »
« less
- NSF-PAR ID:
- 10417643
- Date Published:
- Journal Name:
- 2023 IEEE International Solid- State Circuits Conference (ISSCC)
- Page Range / eLocation ID:
- 342 to 344
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Traditionally, multi-tenant cloud and edge platforms use fair-share schedulers to fairly multiplex resources across applications. These schedulers ensure applications receive processing time proportional to a configurable share of the total time. Unfortunately, enforcing time-fairness across applications often violates energy-fairness, such that some applications consume more than their fair share of energy. This occurs because applications either do not fully utilize their resources or operate at a reduced frequency/voltage during their time-slice. The problem is particularly acute for machine learning (ML) applications using GPUs, where model size largely dictates utilization and energy usage. Enforcing energy-fairness is also important since energy is a costly and limited resource. For example, in cloud platforms, energy dominates operating costs and is limited by the power delivery infrastructure, while in edge platforms, energy is often scarce and limited by energy harvesting and battery constraints. To address the problem, we define the notion of Energy-Time Fairness (ETF), which enables a configurable tradeoff between energy and time fairness, and then design a scheduler that enforces it. We show that ETF satisfies many well-accepted fairness properties. ETF and the new tradeoff it offers are important, as some applications, especially ML models, are time/latency-sensitive and others are energy-sensitive. Thus, while enforcing pure energy-fairness starves time/latency-sensitive applications (of time) and enforcing pure time-fairness starves energy-sensitive applications (of energy), ETF is able to mind the gap between the two. We implement an ETF scheduler, and show that it improves fairness by up to 2x, incentivizes energy efficiency, and exposes a configurable knob to operate between energy- and time-fairness.more » « less
-
We propose a novel method for approximate hardware implementation of univariate math functions with significantly fewer hardware resources compared to previous approaches. Examples of such functions include exp(x) and the activation function GELU(x), both used in transformer networks, gamma(x), which is used in image processing, and other functions such as tanh(x), cosh(x), sq(x), and sqrt(x). The method builds on previous works on hybrid binary-unary computing. The novelty in our approach is that we break a function into a number of sub-functions such that implementing each sub-function becomes cheap, and converting the output of the sub-functions to binary becomes almost trivial. Our method also uses self-similarity in functions to further reduce the cost. We compare our method to the conventional binary, previous stochastic computing, and hybrid binary-unary methods on several functions at 8-, 12-, and 16-bit resolutions. While preserving high accuracy, our method outperforms previous works in terms of hardware cost, e.g., tolerating less than 0.01 mean absolute error, our method reduces the (area x latency) cost on average by 5, 7, and 2 orders of magnitude, compared to the conventional binary, stochastic computing, and hybrid binary-unary methods, respectively. Ultimately, we demonstrate the potential benefits of our method for natural language processing and image processing applications. We deploy our method to implement major blocks in an encoding layer of BERT language model, and also the Roberts Cross edge detection algorithm. Both include non-linear functions.more » « less
-
With the increase in the computation intensity of the chip, the mismatch between computation layer shapes and the available computation resource significantly limits the utilization of the chip. Driven by this observation, prior works discuss spatial accelerators or dataflow architecture to maximize the throughput. However, using spatial accelerators could potentially increase the execution latency. In this work, we first systematically investigate two execution models: (1) sequentially (temporally) launch one monolithic accelerator, and (2) spatially launch multiple accelerators. From the observations, we find that there is a latency throughput tradeoff between these two execution models, and combining these two strategies together can give us a more efficient latency throughput Pareto front. To achieve this, we propose spatial sequential architecture (SSR) and SSR design automation framework to explore both strategies together when deploying deep learning inference. We use the 7nm AMD Versal ACAP VCK190 board to implement SSR accelerators for four end-to-end transformer-based deep learning models. SSR achieves average throughput gains of 2.53x, 35.71x, and 14.20x under different batch sizes compared to the 8nm Nvidia GPU A10G, 16nm AMD FPGAs ZCU102, and U250. The average energy efficiency gains are 8.51x, 6.75x, and 21.22x, respectively. Compared with the sequential-only solution and spatial-only solution on VCK190, our spatial-sequential-hybrid solutions achieve higher throughput under the same latency requirement and lower latency under the same throughput requirement. We also use SSR analytical models to demonstrate how to use SSR to optimize solutions on other computing platforms, e.g., 14nm Intel Stratix 10 NX.more » « less
-
This paper presents ssLOTR (self-supervised learning on the rings), a system that shows the feasibility of designing self-supervised learning based techniques for 3D finger motion tracking using a custom-designed wearable inertial measurement unit (IMU) sensor with a minimal overhead of labeled training data. Ubiquitous finger motion tracking enables a number of applications in augmented and virtual reality, sign language recognition, rehabilitation healthcare, sports analytics, etc. However, unlike vision, there are no large-scale training datasets for developing robust machine learning (ML) models on wearable devices. ssLOTR designs ML models based on data augmentation and self-supervised learning to first extract efficient representations from raw IMU data without the need for any training labels. The extracted representations are further trained with small-scale labeled training data. In comparison to fully supervised learning, we show that only 15% of labeled training data is sufficient with self-supervised learning to achieve similar accuracy. Our sensor device is designed using a two-layer printed circuit board (PCB) to minimize the footprint and uses a combination of Polylactic acid (PLA) and Thermoplastic polyurethane (TPU) as housing materials for sturdiness and flexibility. It incorporates a system-on-chip (SoC) microcontroller with integrated WiFi/Bluetooth Low Energy (BLE) modules for real-time wireless communication, portability, and ubiquity. In contrast to gloves, our device is worn like rings on fingers, and therefore, does not impede dexterous finger motion. Extensive evaluation with 12 users depicts a 3D joint angle tracking accuracy of 9.07° (joint position accuracy of 6.55mm) with robustness to natural variation in sensor positions, wrist motion, etc, with low overhead in latency and power consumption on embedded platforms.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- https://doi.org/10.1109/ISSCC42615.2023.10067817
