This document discusses machine learning building blocks and developing a comprehensive machine learning workload suite. It proposes that there are a finite set of machine learning building blocks that can be mapped to hardware, software, and libraries. These building blocks include linear algebra, measures, special functions, mathematical optimization, data characteristics, data-dependent compute, and memory access. A machine learning workload suite should cover various combinations of these building blocks, algorithms, and relevant datasets to test different compute patterns and data characteristics. The goal is to develop a set of reference workloads to help evaluate machine learning performance on different hardware platforms.
DSPy a system for AI to Write Prompts and Do Fine Tuning
Data! Data! Analytics Building Blocks
1. “Data! Data! Data!
I Can’t Make Bricks Without
Clay!”*
Shai Fine
Principal Engineer, Advanced Analytics, Intel
(*) Sherlock Holmes, The Adventure in the Copper Beeches
4. Analytics to the Rescue
• “Without big data analytics, companies are blind and deaf, wandering
out onto the web like deer on a freeway”
• Geoffrey Moore, Author of Crossing the Chasm
• … and who will lead the way?!
Big Data's High-Priests of Algorithms
The Wall Street Journal, Aug. 2014
5. Adoption of Analytics Faces Hurdles
• Developing Analytics solutions
• Far from being an engineering process
• There is a chasm to cross between “traditional” BI and Advanced Analytics
• Consumability of Analytics
• Deploying Analytics solutions is difficult
• Reliability, “Self Maintenance”
• Analytics Workloads are Challenging
• Speed (latency, time-to-solution), Throughput, Scalability, …
6. The ML Building Blocks Concept
There are “infinite” number of algorithms and datasets
But there are finite set of Building Blocks
Building Blocks:
A finite set of elements that can be mapped into HW and SW primitives and patterns
Building Blocks
Usages
High-level
Libraries
Low-level
Libraries
Hardware
Platforms
Xeon
Xeon Phi
Xeon FPGA
Iris Pro Graphics
Xeon Accel.
New ISA
Tier-1
Cloud
HPC
Enterprise
Academia
7. Machine Learning Building Blocks
• ML basic building blocks
1. Linear Algebra
2. Measures
3. Special Functions
4. Mathematical Optimization
5. Data Characteristics
6. Data-dependent Compute
7. Memory Access
8. Very large models
9. Hybrid Methods
• ML Meta building blocks
1. Learning Protocols
2. Learning Phases
3. Algorithmic Flow and Structure
Compute
Data
Compute - Data Interplay
Process
8. Towards a Comprehensive ML Workload Suite
• Workload design should cover elements of
• Compute
• Data Characteristics
• Data – Compute interplay
• Each workload includes
• Multiple data sets x Multiple algorithms
• Coverage of relevant data characteristics
• Coverage of compute patterns
The Building Block concept provides a mean for designing the ML Workload Suite
9. Machine Learning Workloads Suite
Workload Linear
Algebra
Measure
Calc.
Special
Funcs
Math
Optim.
Data
Characteristics
Data-dep.
Compute
Mem.
Access
large
model
Linear Algebra
Sparse
Dense
X X X
Un/Supervised,
Numeric
Data
Dependency
X X X
Un/Supervised,
Num/Cat
X X
Large Models X X X
Un/Supervised,
Numeric
X
Workload Dataset Type Characteristics
Linear Algebra
Clustered Dense, Numeric
Graphs Sparse, Numeric
Data
Dependency
Bio informatics High Dep - Dense/Sparse
Clustered Dense
Text High Dep – Sparse
Manufacturing High Dep – Numeric, Dense
Large Models Images Dense, Numeric
ALGORITHMS
DATASETS
10. Machine Learning Workloads Suite
Workload Linear
Algebra
Measure
Calc.
Special
Funcs
Math
Optim.
Data
Characteristics
Data-dep.
Compute
Mem.
Access
large
model
Linear Algebra
Sparse
Dense
X X X
Un/Supervised,
Numeric
Data
Dependency
X X X
Un/Supervised,
Num/Cat
X X
Large Models X X X
Un/Supervised,
Numeric
X
Workload Dataset Type Characteristics
Linear Algebra
Clustered Dense, Numeric
Graphs Sparse, Numeric
Data
Dependency
Bio informatics High Dep - Dense/Sparse
Clustered Dense
Text High Dep – Sparse
Manufacturing High Dep – Numeric, Dense
Large Models Images Dense, Numeric
ALGORITHMS
DATASETS
ML Bench 1.0
• Algorithm X Data
• Reference Models
• Data Generator
11. The “Dwarfs” Connection
• Phill Collela’s “Seven Dwarfs” (2004) –
• Patterns of computation and communication
that are important for science and engineering
• Berkley’s view (2006) –
• Extended to 13 Dwarfs after examining
the original 7 Dwarfs outside the HPC scope
• US National Research Council’s Committee
“Frontiers in Massive Data Analysis” (2013) –
• Chapter 10: “The Seven Computational Giants of Massive Data Analysis”
• The ML Building Blocks provide a further extension and a different perspective
• Introducing data characteristics and the interplay with compute, communication, memory