Sai Rahul Chalamalasetti
Sai Rahul Chalamalasetti has a total of 21 co-authors. Here are the top co-authors by number of shared publications:
- John Paul Strachan
Peter Grünberg Institute (PGI-14), Forschungszentrum Jülich GmbH, Jülich, Germany
2 shared publications - Martin Foltin
Artificial Intelligence Research Lab (AIRL), Hewlett Packard Labs, Milpitas, CA, USA
2 shared publications - Aayush Ankit
School of Electrical and Computer Engineering, Purdue University, West Lafayette, IN, USA
1 shared publication - Brent Buchanan
HP Labs, 96953 Palo Alto, California United States
1 shared publication - Can Li
Department of Electrical and Electronic Engineering, The University of Hong Kong, Hong Kong SAR, China
1 shared publication - Catherine Elizabeth Graves
Hewlett Packard Labs, Hewlett Packard Enterprise, Milpitas, USA
1 shared publication
Latest Publications
- Top Cited Papers
- 29 May 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computers
- Vol. 69 (8) , 1128-1142
PANTHER: A Programmable Architecture for Neural Network Training Harnessing Energy-Efficient ReRAM
The wde adoption of deep neural networks has been accompanied by ever-increasing energy and performance demands due to the expensive nature of training them. Numerous special-purpose architectures have been proposed to accelerate training: both digital and hybrid digital-analog using ReRAM crossbars. ReRAM-based accelerators have demonstrated the effectiveness of ReRAM crossbars at performing matrix operations that are prevalent in training. However, they suffer from inefficiency due to the use of serial reads and writes for performing the weight gradient and update step. A few works have demonstrated the possibility of performing outer products in crossbars, which can be used to realize the weight gradient and update step without the use of serial reads and writes. However, these works have been limited to low precision operations which are not sufficient for typical training workloads. Moreover, they have been confined to a limited set of training algorithms for fully-connect layers only. To address these limitations, we propose a bit-slicing technique for enhancing the precision of ReRAM-based outer products. We incorporate this technique into a crossbar architecture with three variants catered to different training algorithms. We build an ISA-programmable training accelerator with compiler support to evaluate our design on both convolutional and fully-connected layers.
- 29 May 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computers
- Vol. 69 (8) , 1128-1142
Top Cited PapersPANTHER: A Programmable Architecture for Neural Network Training Harnessing Energy-Efficient ReRAM
The wde adoption of deep neural networks has been accompanied by ever-increasing energy and performance demands due to the expensive nature of training them. Numerous special-purpose architectures have been proposed to accelerate training: both digital and hybrid digital-analog using ReRAM crossbars. ReRAM-based accelerators have demonstrated the effectiveness of ReRAM crossbars at performing matrix operations that are prevalent in training. However, they suffer from inefficiency due to the use of serial reads and writes for performing the weight gradient and update step. A few works have demonstrated the possibility of performing outer products in crossbars, which can be used to realize the weight gradient and update step without the use of serial reads and writes. However, these works have been limited to low precision operations which are not sufficient for typical training workloads. Moreover, they have been confined to a limited set of training algorithms for fully-connect layers only. To address these limitations, we propose a bit-slicing technique for enhancing the precision of ReRAM-based outer products. We incorporate this technique into a crossbar architecture with three variants catered to different training algorithms. We build an ISA-programmable training accelerator with compiler support to evaluate our design on both convolutional and fully-connected layers.
- 26 August 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Nanotechnology
- Vol. 18 (1536125X) , 963-970
Memristor TCAMs Accelerate Regular Expression Matching for Network Intrusion Detection
We propose memristor-based TCAMs (Ternary Content Addressable Memory) circuits to accelerate Regular Expression (RegEx) matching through in memory processing of finite automata. RegEx matching is a key function in network security to find malicious actors. However, RegEx matching latency and power can be incredibly high and current proposals are challenged to perform wire-speed matching for large rulesets. Our approach dramatically decreases operating power, enables high throughput, and the use of nanoscale memristor TCAM circuits (mTCAMs) enables compression techniques to expand rulesets. We fabricated and demonstrated nanoscale memristor TCAM cells. SPICE simulations investigate performance at scale and a mTCAM dynamic power model using 16nm layout parameters demonstrates ~0.2 fJ/bit/search energy for a 36 × 250 mTCAM array. A tiled architecture is proposed to implement a Snort ruleset and assess application performance. Compared to a state-of-the-art FPGA approach (2 Gbps, ~1W), we show × 4 throughput (8 Gbps) at 55% the power (0.55W) without standard TCAM power-saving techniques. Our performance comparison improves further when striding (searching multiple characters at once) is considered, resulting in 47.2 Gbps at 1.2W for our approach compared to 3.9 Gbps at 630mW for strided FPGA NFA, demonstrating a promising path to wire-speed RegEx matching on large scale rulesets.
- 26 August 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Nanotechnology
- Vol. 18 (1536125X) , 963-970
Memristor TCAMs Accelerate Regular Expression Matching for Network Intrusion Detection
We propose memristor-based TCAMs (Ternary Content Addressable Memory) circuits to accelerate Regular Expression (RegEx) matching through in memory processing of finite automata. RegEx matching is a key function in network security to find malicious actors. However, RegEx matching latency and power can be incredibly high and current proposals are challenged to perform wire-speed matching for large rulesets. Our approach dramatically decreases operating power, enables high throughput, and the use of nanoscale memristor TCAM circuits (mTCAMs) enables compression techniques to expand rulesets. We fabricated and demonstrated nanoscale memristor TCAM cells. SPICE simulations investigate performance at scale and a mTCAM dynamic power model using 16nm layout parameters demonstrates ~0.2 fJ/bit/search energy for a 36 × 250 mTCAM array. A tiled architecture is proposed to implement a Snort ruleset and assess application performance. Compared to a state-of-the-art FPGA approach (2 Gbps, ~1W), we show × 4 throughput (8 Gbps) at 55% the power (0.55W) without standard TCAM power-saving techniques. Our performance comparison improves further when striding (searching multiple characters at once) is considered, resulting in 47.2 Gbps at 1.2W for our approach compared to 3.9 Gbps at 630mW for strided FPGA NFA, demonstrating a promising path to wire-speed RegEx matching on large scale rulesets.