Innovating in a Post Moore’s Law World Mark Horowitz EE & CS, Stanford University 1
Mark Horowitz • Yahoo! Professor, Stanford ▶ Electrical Engineering & Computer Science ▶ Ph.D. in EE from Stanford, 1984 ▶ Former EE Chair • Research: Digital systems design ▶ RISC machines - MIPS-X, TORCH ▶ Distributed Shared Memory – FLASH, SMASH ▶ High-speed IO – Rambus ▶ Security – XOM ▶ Computational Photography – Frankencamera ▶ Extremely Efficient Computing – Darkroom, CE 2
IT World is Changing • Moving from technology to application driven ▶ Success is no longer about access to latest technology ▶ It is about finding the right application to address • To understand why, we need to look at history ▶ Why are computers so prevalent? • How to be successful in this new age 3
Moore’s Law Made Gates Cheap 4
Dennard’s Scaling Made Them Fast & Low Energy • The triple play: 1/ 2 ▶ Get more gates, 1/L 2 ▶ Gates get faster, CV/i 3 ▶ Energy per switch CV 2 Dennard, JSSC, pp. 256‐268, Oct. 1974 5
Our Expectation • Cray-1: world’s fastest computer 1976-1982 ▶ 64Mb memory (50ns cycle time) ▶ 40Kb register (6ns cycle time) ▶ ~1 million gates (4/5 input NAND) ▶ 80MHz clock ▶ 115kW • In 45nm (30 years later) ▶ < 3 mm 2 ▶ > 1 GHz ▶ ~ 1 W CRAY‐1 6
Houston, We Have A Problem http://cpudb.stanford.edu/ 7
The Power Limit Watts/mm 2 http://cpudb.stanford.edu/ 8
We Were Greedy 10x too large http://cpudb.stanford.edu/ 9
This Problem Is Not Going Away: P = C * Vdd 2 * f L 0.6 http://cpudb.stanford.edu/ 10
Think About It 11
Stagnation of Multi-Core Processors http://cpudb.stanford.edu/ 12
Technology to the Rescue? 13
Problems w/ Replacing CMOS • Pretty fundamental physics ▶ Avoiding this problem will be hard e‐ • Its capability is pretty amazing ▶ fJ/gate, 10ps delays, 10 9 working devices 14
Catch - 22 Building Computers = Large $ Very Different = High Risk Investment Risk Capital you need 15
The Truth About Innovation • Start by creating new markets 16
It is the End of Scaling, Not Silicon • Silicon will not disappear ▶ It will still be a huge business, but will consolidate ▶ Growth rate is slower, and scaling is slow • Silicon will become like concrete and steel ▶ Basis of a huge industry, critical to everything ▶ But fairly stable and predictable • Will remain the dominate substrate for computing
Have A Shiny Ball, Now What? 18
Cup Holders • Small additions to a complex product – With large perceived value 19
CPU Cup Holders Specialized Hardware A8 20
Consumer Cup Holders 21
Improved Cup Holders (IoT) • Add communication to compute From Bill Curtis Arm 22
Our CMOS Future • Cup holders made for computing devices ▶ Need to optimize energy efficiency for high performance systems ▶ Build specialized hardware for that application • Cup holders made from computing devices ▶ Capability of today’s technology is incredible ▶ Can add computing and communication for nearly $0 ▶ Key questions are what problems need to be solved? 23
What This Means • Computer performance scaling will slow • Computing chips for specific markets will appear ▶ And manufacturing the addition secret sauce won’t cost very much Computing platforms are stabilizing 24
The New Challenge: • Application specific products have smaller markets ▶ Harder to predict what will win; most will fail ▶ Wins on average are smaller • People who have product ideas ▶ Don’t know about hardware, let alone know how to use it • People who know about the technology ▶ Are a special subset of the population ▶ May not be in touch with what great products will be 25
And System Design Is Hard • Every look at a modern SoC “datasheet”? ▶ They are 500+ pages, and many types • And then you have to worry about the OS ▶ And the drivers 26
The Problem: Last of Clarke’s Three Laws • When a distinguished but elderly scientist states that something is possible, he is almost certainly right. When he states that something is impossible, he is probably wrong. • The only way of discovering the limits of the possible is to venture a little way past them into the impossible. • Any sufficiently advanced technology is indistinguishable from magic. E40M Fall 2015 Lecture 1
Remember This Trade-off? Product Risk • Need to reduce cost to play ▶ Building constructors, not instances Personal cost (time/money) 28
Needed Infrastructure • Apps developers need to work in their space ▶ Program input; auto generate the hardware and system software • Hardware prototypes shipped ▶ Knowledge of fabrication sources ▶ Debugging / bring up support • Sales channel for finished devices ▶ To encourage more people to spend time creating new apps 29
Tock Operating System • Traditionally, embedded systems assume all code is trusted ►No memory protection ►No privilege levels • IoT is moving towards an application store model ►Pebble watch ►iWatch • Need an embedded operating system that supports running multiple, untrusted applications 30
Ravel Framework • Write a data processing pipeline ►Consists of a set of Models , describing data as it is stored ► Transforms move data between Models ►Instances of Models are bound to devices ► Views can display Models ► Controllers determine how data moves to Transforms 31
Key to Success System needs to appeal to two sets of users • Application designers who want to use the system ▶ Need the system to be able to handle many details for them • Expert designers who want to extend the system ▶ Would like it to be “simple” to add new stuff 32
Recently Things Are Looking Up 33
A New Hope • If killer products are going to be application driven ▶ Application experts need to design them • If technology is scaling more slowly ▶ We can incorporate current design knowledge into tools ▶ To create extensible system constructors • We can leverage the 2 nd bullet to enable the 1 st ▶ To usher in a new wave of innovative computing products 34
Recommend
More recommend
