Engineering topics

Somani, Arun K.

Publications and source records attributed to Somani, Arun K..

6 records

DOE OSTI2023

Differentiable Neural Architecture, Mixed Precision and Accelerator Co-Search

Quantization, effective Neural Network architecture, and efficient accelerator hardware are three important design paradigms to maximize accuracy and efficiency. Mixed Precision Quantization is a process of assigning different precision to different Neural Network layers for optimized inference. Neural Architecture Search (NAS) is a process of automatically designing the neural network for a task and can also be extended to search for the precision of each weight and activation matrix. In this paper, we develop the following three methods: (i) Fast Differentiable Hardware-aware Mixed Precision Quantization Search method to find optimal precision, (ii) Joint Differentiable hardware-aware Architecture and Mixed Precision Quantization Co-search, (iii) Joint Accelerator, Architecture, and Precision triple co-search to find best possibilities in all the three worlds. We demonstrate the effectiveness of our proposed methods targeting Bitfusion accelerator by searching mixed precision models on MobilenetV2. We achieve better accuracy-latency trade-off models than the manually designed and previously proposed search methods.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2023

A survey of techniques for optimizing transformer inference

Recent years have seen a phenomenal rise in the performance and applications of transformer neural networks. The family of transformer networks, including Bidirectional Encoder Representations from Transformer (BERT), Generative Pretrained Transformer (GPT) and Vision Transformer (ViT), have shown their effectiveness across Natural Language Processing (NLP) and Computer Vision (CV) domains. Transformer-based networks such as ChatGPT have impacted the lives of common men. However, the quest for high predictive performance has led to an exponential increase in transformers' memory and compute footprint. Researchers have proposed techniques to optimize transformer inference at all levels of abstraction. Further, this paper presents a comprehensive survey of techniques for optimizing the inference phase of transformer networks. We survey techniques such as knowledge distillation, pruning, quantization, neural architecture search and lightweight network design at the algorithmic level. We further review hardware-level optimization techniques and the design of novel hardware accelerators for transformers. We summarize the quantitative results on the number of parameters/FLOPs and the accuracy of several models/techniques to showcase the tradeoff exercised by them. We also outline future directions in this rapidly evolving field of research. We believe that this survey will educate both novice and seasoned researchers and also spark a plethora of research efforts in this field.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2023

Neural Architecture Search Benchmarks: Insights and Survey

Neural Architecture Search (NAS), a promising and fast-moving research field, aims to automate the architectural design of Deep Neural Networks (DNNs) to achieve better performance on the given task and dataset. NAS methods have been very successful in discovering efficient models for various Computer Vision, Natural Language Processing, etc. The major obstacles to the advancement of NAS techniques are the demand for large computation resources and fair evaluation of various search methods. The differences in training pipeline and setting make it challenging to compare the efficiency of two NAS algorithms. A large number of NAS Benchmarks to simulate the architecture evaluation in seconds have been released over the last few years to ease the computation burden of training neural networks and can aid in the unbiased assessment of different search methods. This paper provides an extensive review of several publicly available NAS Benchmarks in the literature. We provide technical details and a deeper understanding of each benchmark and point out future directions.

97 MATHEMATICS AND COMPUTING↗

NASA NTRS1996

Simplified Phased-Mission System Analysis for Systems with Independent Component Repairs

Accurate analysis of reliability of system requires that it accounts for all major variations in system's operation. Most reliability analyses assume that the system configuration, success criteria, and component behavior remain the same. However, multiple phases are natural. We present a new computationally efficient technique for analysis of phased-mission systems where the operational states of a system can be described by combinations of components states (such as fault trees or assertions). Moreover, individual components may be repaired, if failed, as part of system operation but repairs are independent of the system state. For repairable systems Markov analysis techniques are used but they suffer from state space explosion. That limits the size of system that can be analyzed and it is expensive in computation. We avoid the state space explosion. The phase algebra is used to account for the effects of variable configurations, repairs, and success criteria from phase to phase. Our technique yields exact (as opposed to approximate) results. We demonstrate our technique by means of several examples and present numerical results to show the effects of phases and repairs on the system reliability/availability.

Somani, Arun K.↗

NASA NTRS1995

Architecture and performance analysis of DIRSMIN: A fault-tolerant switch using dilated reduced-stage MIN

We develop and analyze a dilated high performance fault tolerant fast packet multistage interconnection network (MIN) in this paper. In this new design, the links at the input and the output stages of a dilated banyan-based MIN are rearranged to create multiple routes for each source-destination pair in the network after removing one stage in the network. These multiple paths are link- and node-disjoint. Fault tolerance at low latency is achieved by sending multiple copies of each input packet simultaneously using different routes and different priorities. This guarantees that high throughput is maintained even in the presence of faults. Throughput is analyzed using simulation and analysis and we show that the new design has considerably higher performance in the presence of a faulty switching element (SE) or link in comparison to dilated networks. We also analyze the reliability and show that the new design has superior reliability in comparison to competing proposals.

Somani, Arun K.↗

NASA NTRS1993

Phased-mission system analysis using Boolean algebraic methods

Most reliability analysis techniques and tools assume that a system is used for a mission consisting of a single phase. However, multiple phases are natural in many missions. The failure rates of components, system configuration, and success criteria may vary from phase to phase. In addition, the duration of a phase may be deterministic or random. Recently, several researchers have addressed the problem of reliability analysis of such systems using a variety of methods. A new technique for phased-mission system reliability analysis based on Boolean algebraic methods is described. Our technique is computationally efficient and is applicable to a large class of systems for which the failure criterion in each phase can be expressed as a fault tree (or an equivalent representation). Our technique avoids state space explosion that commonly plague Markov chain-based analysis. A phase algebra to account for the effects of variable configurations and success criteria from phase to phase was developed. Our technique yields exact (as opposed to approximate) results. The use of our technique was demonstrated by means of an example and present numerical results to show the effects of mission phases on the system reliability.

Somani, Arun K.↗