CogniGron seminar-Computer Science - Prof. Luigi Carro, UFRGS, Brazil

Title: Accelerating ChatGPT and other complex NN using ReRAM devices

Abstract:
As the massive usage of Artificial Intelligence (AI) techniques
spreads in the economy, researchers are exploring new techniques to
reduce the energy consumption of Neural Network (NN) applications,
especially as the complexity of NNs continues to increase. Using
analog Resistive RAM (ReRAM) devices to compute Matrix-Vector
Multiplication (MVM) in (1) time complexity is a promising approach,
but it’s true that these implementations often fail to cover the
diversity of nonlinearities required for modern NN applications, that
reach 37% of computing costs. In this presentation we propose a novel
approach where ReRAMs themselves can be reprogrammed to compute not
only the required matrix multiplications, but also the activation
functions, pooling, and tokenization layers, reducing energy in
complex NNs. We discuss our results in experiments on real-world human
activity recognition and language modeling datasets with Convolutional
Neural Networks (CNNs), Generative Pre-trained Transformer (GPT), and
Long Short-Term Memory (LSTM) models, where we compare our strategy
with different platforms, from GPUs to FPGAs.

Short bio:

Luigi Carro received the Electrical Engineering and the MSc degrees from Universidade Federal do Rio Grande do Sul (UFRGS), Brazil, in 1985 and 1989, respectively. From 1989 to 1991 he worked at ST-Microelectronics, Agrate, Italy, in the R&D group. In 1996 he received the Dr. degree in the area of Computer Science from Universidade Federal do Rio Grande do Sul (UFRGS), Brazil. He is presently a full professor at the Applied Informatics Department at the Informatics Institute of UFRGS, in charge of Computer Architecture and Organization. He has advised more than 20 graduate students, and has published more than 150 technical papers on those topics. He has authored the book Digital systems Design and Prototyping (2001-in Portuguese) and is the co-author of Fault-Tolerance Techniques for SRAM-based FPGAs (2006-Springer), Dynamic Reconfigurable Architectures and Transparent optimization Techniques (2010-Springer) and Adaptive Systems (Springer 2012). In 2007 he received the prize APERGS - Researcher of the year in Computer Science. https://www.inf.ufrgs.br/~carro/

Online to follow via: https://bluejeans.com/636348579/1898