ML/Hardware Co-design: Overview, Preliminary Result, and Open Opportunities Ce Zhang

Transcription

1 ML/Hardware Co-design: Overview, Preliminary Result, and Open Opportunities Ce Zhang

2 Machine Learning: Why should we care?

3 plus some other (equally important) reasons! 3

4 4 Machine Learning ETH (Some Samples)

5 ONE SIMPLE REASON If people who have 20 Science&Nature papers think machine learning is important to treat us when we are sick, we d better help them with that.

6 Overview of This Lecture 6 1. (Whiteboard - 20 mins) Overview of Machine Learning from a System Perspective - How many of you know Linear Regression? - How many of you know Support Vector Machine? - How many of you know Gradient Descent? - How many of you know Stochastic Gradient Descent? 2. (15 mins) A sample of our previous work related to hardware & machine learning - NUMA - CPU vs. GPU 3. (10 mins) Ongoing - - Low Precision Arithmetic

7 A Short Tutorial: Linear Regression Patient Survival Time Predict

8 A Short Tutorial: Linear Regression Patient Survival Time Features Age Weight Stage Predict

9 A Short Tutorial: Linear Regression Patient Survival Time Features Age Weight Stage Predict Data Age Weight Stage Days ? Training Testing

10 A Short Tutorial: Linear Regression Data Goal of Training Millions/Billions! Age Heig Stag A Days ? b Loss (Sum of prediction errors) Model Age Heig Stag x Assumption The prediction of r th patient is! " $ r th row of A

11 More General than Linear Regression Many machine learning tasks can be written as - min % &((, * + ) $ +./ Billions! 28 28

12 Whiteboard 1.Gradient Descent 2.Stochastic Gradient Descent 3.Analytical Comparison between SGD and GD 4.Other Smarter ways to design the gradient estimator 1.Mini-batch 2.Low-precision 5.A Taxonomy of System Bottlenecks (Single Worker) 6.Distributed Setting

13 Low Precision Arithmetic

14 Limited Precision 14 Can we do low-precision data (not gradient)? More difficult. (Can you see why?) Whiteboard Single Bit Gradient? NVLINK, DGX-1 Can we do near memory computation here?!

15 15 Linear Regression: 32-bit Data => 1-bit Data /1.5 = 21x Bandwidth Reduction?! Better Training Loss bit 1.5bit Double Sampling (We get this curve :) 1-bit Random Rounding # Iterations

16 Deep Learning: CPU vs. GPU

17 Whiteboard Convolution & Neural Network

18 Single-Machine: CPU vs. GPU On which machine should we run? CPU or GPU? I have a GPU Cluster I have 5000 CPU cores I have $100K to spend on the cloud EC2: c4.4xlarge 8 cores@2.90ghz 0.7TFlops EC2: g2.2xlarge 1.5K cores@800mhz 1.2TFlops Not a 10x gap? Can we close this gap?

19 Caffe con Troll A prototype system to study the CPU/GPU tradeoff. Same-input-same-output as Caffe.

20 What we found 1.1 Relative Speed Caffe CPU CcT CPU Caffe GPU Caffe CPU CcT CPU c4.4x_large ($0.68/h) c4.4x_large ($0.68/h) g2.2x_large ($0.47/h) c4.8x_large ($1.37/h) c4.8x_large ($1.37/h)

21 What we found 1.1 Relative Speed Caffe CPU CcT CPU Caffe GPU Caffe CPU CcT CPU c4.4x_large ($0.68/h) c4.4x_large ($0.68/h) g2.2x_large ($0.47/h) c4.8x_large ($1.37/h) c4.8x_large ($1.37/h)

22 What we found 1.1 Relative Speed Caffe CPU CcT CPU Caffe GPU Caffe CPU CcT CPU c4.4x_large ($0.68/h) c4.4x_large ($0.68/h) g2.2x_large ($0.47/h) c4.8x_large ($1.37/h) c4.8x_large ($1.37/h)

23 What we found 1.1 Relative Speed Proportional to FLOPs! Caffe CPU CcT CPU Caffe GPU Caffe CPU CcT CPU c4.4x_large ($0.68/h) c4.4x_large ($0.68/h) g2.2x_large ($0.47/h) c4.8x_large ($1.37/h) c4.8x_large ($1.37/h)

24 Four Shallow Ideas Described in Four Pages arxiv:

25 One of the four shallow ideas 3 CPU Cores 3 Images Strategy 1 Strategy 2 If the amount of data is too small for each core, the process might not be CPU bound. For AlexNet over Haswell CPUs, Strategy 2 is 3-4x faster.

26 Ongoing Work

27 Where to run Machine Learning?

28 Big Data in Small Pockets: A Vision Hypothesis 1: 2min on Jetson TX1 How fast can I train an SVM over 100GB data? PCIe Gen2 4x 6sec = 1Epoch Jetson TX1 1TFLOPS, 4GB RAM 15W FPGA Data Compression? Phone Energy Efficiency? Go Beyond Deep Learning How fast can I train SVM+LR over 100GB? Hypothesis 2: Same as 1 SVM (Bandwidth Bound) Batching? Model Compression (Hashing Trick for Training)? Which operations should happen on host/device?

29 Precision-aware Network Protocol

30 How Should Limited Precision Primitives be Supported? Low-precision Network Primitives? TCP/IP for ML? 30 TensorFlow CNTK Caffe MxNET SINGA Precision Aware Network Protocol Network Need FPGA to be fast?

31 How to Program Machine Learning?

32 32 Three Worlds: One System? Machine Learning - Tensor, Tensor, Tensor - Real Algebra Database - Relation - Relational Algebra Matrix Factorization let matrix be float[][] let result be (P, Q) where P is float[][] and R is float[][] COPY matrix from matrix.tsv Partial Schema (Static Analysis to figure out the rest?) SQL/Spark Spark - Tuples - Map Reduce COPY result from `zero()` v_{i,j} = P_{i,-} * Q_{j,-} Real-Algebra View Logically, not different, all DataCube Can we bridge the gap of physical representation? loss = \sum_{i,j}(\sqrt(v_{i,j} matrix_{i,j})) min loss over P and Q Real-Algebra Query LaTeX-compilant (Math should look like Math)

33 How to Guide user to use Machine Learning?

34 Cheat Sheet 34 Are you serious? What does it mean to be a good Cheat sheet? HCI => I want a quantitive measure If X then Y: - How to give Premise? - How to give Consequence?

35 How to take advantage Machine Learning?

36 More Intelligent/Robust Systems 36 What would happen if you ask today's DB System a query that it was not programmed to answer? DB System Given q, Find d: min q /\ d / d Given q, Find d: min q /\ d / q \/ d WTF!@#%#@ I don t understand!! SMT Solver to automatically figure out how to answer new queries Simple, but we made the production system of a leading security company 100x faster

37 Even Crazier 37 Can we beat human in generating Assembly? Not a new dream Super optimizer (Henry Massalin 1987) Beat Human in Go! BUT we are at the right time to making it come true!

38 How to change the world with Machine Learning?

39 Applications Cyber security: FireEye & IBM Watson Astrophysics: ETH Patient Stratification: UZH Hospital & ETH Genomics: ETH Aneurysm: ZHAW Social Science: ETH Dialogue System: BMW More to come! 39 How to support 100 applications with sublinear cost (development time, machine time, management time)? Our final goal in the next K years.