Link Search Menu Expand Document (external link)

Chips & Compilers Symposium at MLSys ‘22

Modern machine learning systems call for scalable and efficient solutions to both gigantic model training and flexible model inference. This requires joint design and optimization between hardware and software to take full advantage of the provided resource. We have observed active development in this field in the past few years, ranging from machine learning specific hardware and compiler, to customized optimization towards modern machine learning workloads. This symposium aims to bring together experts from the field of computer architecture and compilers to share first-hand experiences, lessons, and best practices, of designing ML workload specific chips and compilers. Like last year, the symposium consists of invited talks by domain experts from both academia and industry.

Time and location

Mission Ballroom MR3, Santa Clara Convention Center

9 am, September 1, 2022

Speakers

Dave Patterson

Distinguished Engineer at Google

Fredrik Kjolstad

Assistant Professor at Stanford

Hadi Esmaeilzadeh

Associate Professor at UCSD and Co-Founder and CTO at Protopia AI

Jianhui Li

Senior Principal Engineer at Intel

Peng Wu

Engineering Manager at Meta

Ron Diamant

Senior Principal Engineer at AWS

Song Han

Associate Professor at MIT

Tobias Edler von Koch

Senior Compiler Engineer at AWS

Vijay Janapa Reddi

Associate Professor at Harvard

Yuanzhong Xu

Staff Software Engineer at Google

Zhihao Jia

Assistant Professor at CMU

Organizers

Yida Wang

Principal Scientist at AWS

Gennady Pekhimenko

Assistant Professor at University of Toronto and Vector Institute

Agenda

Morning Session

Time
Talk Title
Speaker
9:00–9:05
Opening
Organizers
9:05–9:50
A Decade of Machine Learning Accelerators: Lessons Learned and Carbon Footprint
Dave Patterson
9:50–10:20
Automatically Discovering Machine Learning Optimizations
Zhihao Jia
10:20–10:40

Break

10:40–11:10
What makes PyTorch beloved makes it hard to compile – the nuanced design of PyTorch Compilers
Peng Wu
11:10–11:40
Tiny Machine Learning
Vijay Janapa Reddi
11:40–12:10
AI for Better AI
Hadi Esmaeilzadeh
12:10–13:15

Lunch (Will be provided)

Afternoon Session

Time
Talk Title
Speaker
13:15–14:00
Optimizing ML workloads across the stack with AWS Trainium
Ron Diamant & Tobias Edler von Koch
14:00–14:30
GSPMD: generalized parallelism for large models as shared compiler infrastructure
Yuanzhong Xu
14:30–15:00

Break

15:00–15:30
Software and Hardware for Sparse ML
Fred Kjolstad
15:30–16:00
On-Device Training Under 256KB Memory
Song Han
16:00–16:30
Generating efficient code for deep neural network
Jianhui Li

See also