Lecture

Deep Learning Accelerators: Optimization Strategies

Graph Chat

Chat with Graph Search

Ask any question about EPFL courses, lectures, exercises, research, news, etc. or try the example questions below.

DISCLAIMER: The Graph Chatbot is not programmed to provide explicit or categorical answers to your questions. Rather, it transforms your questions into API requests that are distributed across the various IT services officially administered by EPFL. Its purpose is solely to collect and recommend relevant references to content that you can explore to help you answer your questions.

In course

CS-471: Advanced multiprocessor architecture

Multiprocessors are basic building blocks for all computer systems. This course covers the architecture and organization of modern multiprocessors, prevalent accelerators (e.g., GPU, TPU), and datacenters. It includes a research project on multiprocessors ...

Description

This lecture delves into the optimization strategies for deep learning accelerators, focusing on reducing data movement in DNNs through techniques such as batching, dataflow optimizations, and compression. The instructor explains the trade-offs involved in batching, the benefits of dataflow optimizations for energy efficiency, and the challenges and potential of compression techniques. Various concepts like weight reuse, input reuse, and different dataflow architectures are explored to maximize the efficiency of deep learning accelerators.

Instructor

Babak Falsafi

Babak is a Professor in the School of Computer and Communication Sciences and the founding director of the EcoCloud, an industrial/academic consortium at EPFL investigating scalable data-centric technologies. He has made numerous contributions to computer system design and evaluation including a scalable multiprocessor architecture which was prototyped by Sun Microsystems (now Oracle), snoop filters and memory streaming technologies that are incorporated into IBM BlueGene/P and Q and ARM cores, and computer system performance evaluation methodologies that have been in use by AMD, HP and Google PerKit . He has shown that hardware memory consistency models are neither necessary (in the 90's) nor sufficient (a decade later) to achieve high performance in multiprocessor systems. These results eventually led to fence speculation in modern microprocessors. His latest work on workload-optimized server processors laid the foundation for the first generation of Cavium ARM server CPUs, ThunderX. He is a recipient of an NSF CAREER award, IBM Faculty Partnership Awards, and an Alfred P. Sloan Research Fellowship. He is a fellow of IEEE and ACM.

Official source

Related concepts (32)

Data compression

In information theory, data compression, source coding, or bit-rate reduction is the process of encoding information using fewer bits than the original representation. Any particular compression is either lossy or lossless. Lossless compression reduces bits by identifying and eliminating statistical redundancy. No information is lost in lossless compression. Lossy compression reduces bits by removing unnecessary or less important information.

Lossy compression

In information technology, lossy compression or irreversible compression is the class of data compression methods that uses inexact approximations and partial data discarding to represent the content. These techniques are used to reduce data size for storing, handling, and transmitting content. The different versions of the photo of the cat on this page show how higher degrees of approximation create coarser images as more details are removed. This is opposed to lossless data compression (reversible data compression) which does not degrade the data.

Lossless compression

Lossless compression is a class of data compression that allows the original data to be perfectly reconstructed from the compressed data with no loss of information. Lossless compression is possible because most real-world data exhibits statistical redundancy. By contrast, lossy compression permits reconstruction only of an approximation of the original data, though usually with greatly improved compression rates (and therefore reduced media sizes).

Compression artifact

A compression artifact (or artefact) is a noticeable distortion of media (including , audio, and video) caused by the application of lossy compression. Lossy data compression involves discarding some of the media's data so that it becomes small enough to be stored within the desired or transmitted (streamed) within the available bandwidth (known as the data rate or bit rate). If the compressor cannot store enough data in the compressed version, the result is a loss of quality, or introduction of artifacts.

Image compression

Image compression is a type of data compression applied to s, to reduce their cost for storage or transmission. Algorithms may take advantage of visual perception and the statistical properties of image data to provide superior results compared with generic data compression methods which are used for other digital data. Image compression may be lossy or lossless. Lossless compression is preferred for archival purposes and often for medical imaging, technical drawings, clip art, or comics.

Related lectures (64)

Compression: Prefix-Free Codes

Explains prefix-free codes for efficient data compression and the significance of uniquely decodable codes.

Data Compression and Shannon's Theorem: Lossy Compression

Explores data compression, including lossless methods and the necessity of lossy compression for real numbers and signals.

Neural Signal Compression

Explores analog-to-digital conversion, neural signal optimization, multichannel architectures, and on-chip compression techniques in neuroengineering.

Image compression

Explores image compression through various approaches like pixel and block level compression, Discrete Cosine Transform, quantization, and entropy coding.

Second-Order Model Compression

Explores second-order model compression for massive deep neural networks, showcasing compression techniques and their impact on model accuracy.

Related concepts (32)