Jun 14, 2015 - COMPUTER ARCHITECTURE FOR ... Machine Learning Workloads Analysis. 11:30-12:00 ... Accelerators for Machi
COMPUTER ARCHITECTURE FOR MACHINE LEARNING ISCA-42
Portland OR, USA
June 14, 2015
Organizers: Boris Ginsburg, Ronny Ronen - Intel Labs Olivier Temam - Google
Program Time
Speaker
Title
09:00-9:20
Boris Ginsburg, Intel
Opening Remarks
Hardware Acceleration for Deep Learning 09:20-9:45
Amir Khosrowshahi, Nervana Systems 9:45-10:10 Eric Chung, Microsoft Research 10:10-10:35 Vinayak Gokhale, Purdue University 10:35-11:00 Paul Burchard Goldman Sachs
Computer Architectures for Deep Learning Accelerating Deep Convolutional Neural Networks Using Specialized Hardware in the Datacenter A Hardware Accelerator for Convolutional Neural Networks Hardware Acceleration for Communication-Intensive Algorithms
11:00-11:30 COFFE
Machine Learning Workloads Analysis 11:30-12:00 Jonathan Pearce, Intel Labs 12:00-12:30 Scott Beamer, UC Berkeley, 12:30–13:30 LUNCH
You Have No (Predictive) Power Here, SPEC! Graph Processing Bottlenecks
Program Time
Speaker
Title Neuromorphic Engineering
13:30-14:00 James E. Smith Wisconsin–Madison 14:00-14:30 Giacomo Indiveri, Univ. of Zurich and ETH Zurich 14:30-15:00 Yiran Chen, Univ. of Pittsburgh 15:00-15:30 Mikko Lipasti, Wisconsin – Madison
Biologically Plausible Spiking Neural Networks Neuromorphic circuits for building autonomous cognitive systems Hardware Acceleration for Neuromorphic Computing: An Evolving View Mimicking the Self-Organizing Properties of the Visual Cortex
15:30-16:00 COFFE
Hardware Acceleration for Machine Learning 16:00-16:30 Shai Fine, Intel 16:30-17:00 Chunkun Bo, University of Virginia 17:00-17:30 Ran Ginosar, Technion
Machine Learning Building Blocks String Kernel Testing Acceleration using the Micron Automata Processor Accelerators for Machine Learning of Big Data
Paradigm Shift: From Computers to Learning Machines We are in the very beginning of new computer era
Traditional Computer Architecture 1946 - ENIAC Electronic Numerical Integrator And Computer, designed by J. P. Eckert and J. Mauchly
First Programming Model ENIAC was programmed by setting switches and inserting patch leads to route data and to control signals between various functional units
Von Neumann Architecture EDVAC - stored-program computer: • CPU with ALU , Control unit, and registers • Memory to store data and program • Synchronous
1948-2015: Same Old Programming Model
Programs = Algorithms + Data Structures • Algorithms are translated into linear sequence of instructions, which are executed synchronously • Data Structures are mapped to Linear Memory
Paradigm Shift: From Formulas and Algorithms to Machine Learning
Machine Learning – breakthrough in computer vision, natural language processing, speech recognition, robotics, self-driving cars,…
Machine Learning Building Blocks Machine Learning Convolutional NN
GMM
HMM
Recurrent NN
SVM
DBN
Auto-Encoders
kNN
MLP
Dense / Sparse Matrix
Graph Algorithms
?
Dense / Sparse BLAS & FFT
“Think like a Vertex”
?
Graph Accelerator
?
Matrix Co-processor 10
Deep Learning – Just First Step… New way to develop application: • explicit algorithms and formulas • multi-layer neural networks trained on large data
Paradigm Shift: Neuromorphic Engineering Traditional Computers • • • • •
Linear execution model Flat linear memory model Synchronous Numerically precise Reliable
Neuromorphic • • • • •
Parallel execution Compute-in-memory Asynchronous Probability computing unreliable elements
What it takes to make this paradigm shift? A lot of things… Including: • Machine Learning Workloads Analysis • Hardware Acceleration for Machine Learning • Hardware Acceleration for Deep Learning • Neuromorphic Engineering
In short – our program…
