[Workshop] ScaledML: Moore's Law in the age of AI Chips
Moore's Law in the age of AI Chips1
The workshop was presented by Jim Keller, Intel.
Moore's Law
Moore ’ S Law Transistors
Brought by both new architectures and new materials.
Transistor Scaling
Paradigm Shift
1000 Scalars
Complexity Limits
Why not big computers?
Key Problems:
- How to program 1 million computers?
- Address > 260 objects?
- Make 1 peta look-ups into 1 petabyte of data?
Abstraction Layers
Compute models
Data models
GPUs & AI Chips
GPUs
AI Chips
- SIMD
- Local Memory
- No privilege model
- hard to program
Technology Optimism
Summary
-
ScaledML, http://scaledml.org/2020/ ↩