cse240a: Graduate Computer Architecture Steven Swanson Hung-Wei - PowerPoint PPT Presentation

cse240a: Graduate Computer Architecture Steven Swanson Hung-Wei Tseng 1

Today’s Agenda • What is architecture? • Why is it important? • At the highest level, where is architecture today? Where is it going? • What’s in this class? 2

What is architecture? • How do you build a machine that computes? • Quickly, safely, cheaply, efficiently, in technology X, for application Y, etc. Civilization advances by extending the number of important operations which we can perform without thinking about them. -- Alfred North Whitehead

Orientation The internet

Orientation The internet

Orientation System Bus IO (PCI) Memory Memory Power Power Memory Memory Architecture begins about here.

Orientation System Bus IO (PCI) Memory Memory Power Power Memory Memory Architecture begins about here.

You are here

You are here cse240a

The processors go here…

The processors go here…

Abstractions of the Physical World… Physics/Materials Devices Micro-architecture Processors Architectures

Abstractions of the Physical World… cse241a/ Physics/ This Course ECE dept Chemistry/ Material science Physics/Materials Devices Micro-architecture Processors Architectures

…for the Rest of the System JVM Processor Software Compilers Languages Architectures Abstraction Engineers/ Applications

…for the Rest of the System JVM Processor Software Compilers Languages Architectures Abstraction Engineers/ Applications

Why study architecture? • As CEs or CSs you should understand how computers work • Processors are the basis for everything in CS (except theory) • They are where the rubber meets the road. • Performance is important • Faster machines make applications cheaper • Understanding hardware is essential to understanding how systems behave • It’s cool! • Microprocessors are among the most sophisticated devices manufactured by people • How they work (and even that they work) as reliably and as quickly as they do is amazing. • Architecture is undergoing a revolution • The future is uncertain • Opportunities for innovation abound. 11

Performance and You! • Live Demo 12

Processor are Cool! • Chips are made of silicon • Aka “sand” • The most abundant element in the earth’s crust. • Extremely pure (<1 part per billion) • This is the purest stuff people make

Building Chips

Building Chips • Photolithography Silicon Wafer

Building Chips • Photolithography SiO2 Silicon Wafer Silicon Wafer Grow silicon dioxide

Building Chips • Photolithography Resist SiO2 SiO2 Silicon Wafer Silicon Wafer Silicon Wafer Grow silicon dioxide Apply photo resist

Building Chips • Photolithography Mask Mask Resist Resist SiO2 SiO2 SiO2 Silicon Wafer Silicon Wafer Silicon Wafer Silicon Wafer Grow silicon dioxide Apply photo resist Expose to UV

Building Chips • Photolithography Mask Mask Resist Resist SiO2 SiO2 SiO2 Silicon Wafer Silicon Wafer Silicon Wafer Silicon Wafer Grow silicon dioxide Apply photo resist Expose to UV SiO2 Silicon Wafer Patterned resist

Building Chips • Photolithography Mask Mask Resist Resist SiO2 SiO2 SiO2 Silicon Wafer Silicon Wafer Silicon Wafer Silicon Wafer Grow silicon dioxide Apply photo resist Expose to UV SiO2 Silicon Wafer Silicon Wafer Patterned resist Etch SiO2

Building Chips • Photolithography Mask Mask Resist Resist SiO2 SiO2 SiO2 Silicon Wafer Silicon Wafer Silicon Wafer Silicon Wafer Grow silicon dioxide Apply photo resist Expose to UV SiO2 Met Silicon Wafer Silicon Wafer Silicon Wafer Patterned resist Deposit metal Etch SiO2

Building Chips • Photolithography Mask Mask Resist Resist SiO2 SiO2 SiO2 Silicon Wafer Silicon Wafer Silicon Wafer Silicon Wafer Grow silicon dioxide Apply photo resist Expose to UV SiO2 Met Met Silicon Wafer Silicon Wafer Silicon Wafer Silicon Wafer Etch SiO2 Patterned resist Deposit metal Etch SiO2 (Or not)

Building Blocks: Transistors

Building Blocks: Wires

State of the art CPU • 1-2 Billion xtrs • 45nm features • 3-4Ghz • Several 100 designers • >5 years • $3Billion fab • 70 GFLOPS 18

Current state of architecture

Since 1940

Since 1940 • Plug boards -> Java • Hand assembling -> GCC • No OS -> Windows Vista

Since 1940 • Plug boards -> Java • 50,000 x speedup • Hand assembling -> GCC • >1,000,000,000 x density • No OS -> Windows Vista (Moore’s Law) Flexible performance is a liquid asset

Moore’s Law: Raw transistors

Computer Performance 22

Computer Performance 10000 specINT95 specINT2000 specINT2006 1000 Relative Performance 100 10 1 1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010 Year 22

Computer Performance 10000 10000 specINT95 specINT95 specINT2000 specINT2000 specINT2006 specINT2006 47% per year 1000 1000 Relative Performance Relative Performance 100 100 10 10 1 1 1990 1990 1992 1992 1994 1994 1996 1996 1998 1998 2000 2000 2002 2002 2004 2004 2006 2006 2008 2008 2010 2010 Year Year 22

Computer Performance 10000 10000 10000 specINT95 specINT95 specINT95 specINT2000 specINT2000 specINT2000 specINT2006 specINT2006 specINT2006 47% per year 47% per year 39% per year 1000 1000 1000 Relative Performance Relative Performance Relative Performance 100 100 100 10 10 10 1 1 1 1990 1990 1990 1992 1992 1992 1994 1994 1994 1996 1996 1996 1998 1998 1998 2000 2000 2000 2002 2002 2002 2004 2004 2004 2006 2006 2006 2008 2008 2008 2010 2010 2010 Year Year Year 22

Computer Performance 10000 10000 10000 10000 specINT95 specINT95 specINT95 specINT95 specINT2000 specINT2000 specINT2000 specINT2000 specINT2006 specINT2006 specINT2006 specINT2006 47% per year 47% per year 47% per year 39% per year 39% per year 25% per year 1000 1000 1000 1000 Relative Performance Relative Performance Relative Performance Relative Performance 100 100 100 100 10 10 10 10 1 1 1 1 1990 1990 1990 1990 1992 1992 1992 1992 1994 1994 1994 1994 1996 1996 1996 1996 1998 1998 1998 1998 2000 2000 2000 2000 2002 2002 2002 2002 2004 2004 2004 2004 2006 2006 2006 2006 2008 2008 2008 2008 2010 2010 2010 2010 Year Year Year Year 22

The clock speed addiction specINT2000 specINT2006 5000 4000 Clock speed (Mhz) 3000 2000 1000 0 1996 1998 2000 2002 2004 2006 2008 2010 Year • Clock speed is the biggest contributor to power • Chip manufactures (Intel, esp.) pushed clock speeds very hard in the 90s and early 2000s. • Doubling the clock speed increases power by 2-8x • Clock speed scaling is essentially finished. 23

Power 1000 100 2 Watts/cm 10 1 1.5µ 1µ 0.7µ 0.5µ 0.35µ 0.25µ 0.18µ 0.13µ 0.1µ 0.07µ 1.5µ 1µ 0.7µ 0.5µ 0.35µ 0.25µ 0.18µ 0.13µ 0.1µ 0.07µ 24

Power 1000 100 2 Watts/cm 10 1 1.5µ 1µ 0.7µ 0.5µ 0.35µ 0.25µ 0.18µ 0.13µ 0.1µ 0.07µ 1.5µ 1µ 0.7µ 0.5µ 0.35µ 0.25µ 0.18µ 0.13µ 0.1µ 0.07µ 24

Power 1000 100 2 Watts/cm 10 1 1.5µ 1µ 0.7µ 0.5µ 0.35µ 0.25µ 0.18µ 0.13µ 0.1µ 0.07µ 1.5µ 1µ 0.7µ 0.5µ 0.35µ 0.25µ 0.18µ 0.13µ 0.1µ 0.07µ 24

Power 1000 100 2 Watts/cm 10 1 1.5µ 1µ 0.7µ 0.5µ 0.35µ 0.25µ 0.18µ 0.13µ 0.1µ 0.07µ 1.5µ 1µ 0.7µ 0.5µ 0.35µ 0.25µ 0.18µ 0.13µ 0.1µ 0.07µ 24

Power 1000 100 2 Watts/cm 10 1 1.5µ 1µ 0.7µ 0.5µ 0.35µ 0.25µ 0.18µ 0.13µ 0.1µ 0.07µ 1.5µ 1µ 0.7µ 0.5µ 0.35µ 0.25µ 0.18µ 0.13µ 0.1µ 0.07µ 24

Power 1000 100 2 Watts/cm 10 1 1.5µ 1µ 0.7µ 0.5µ 0.35µ 0.25µ 0.18µ 0.13µ 0.1µ 0.07µ 1.5µ 1µ 0.7µ 0.5µ 0.35µ 0.25µ 0.18µ 0.13µ 0.1µ 0.07µ 24

What’s Next: Brainiacs • Hold the clock rate steady. • Be smarter in silicon • More sophisticated processors • More clever algorithms • This continues to deliver about 25% per year. • But for how long? 25

What’s Next: Parallelism • This is all the rage right now • You probably own a multi-processor, they used to be pretty exotic. • They provide some performance, but it’s hard to use. • There aren’t that many threads • Remember, flexible performance is a liquid asset • Remember or look forward to OS 26

Intel P4 Intel Core 2 Duo Intel Nahalem 1 core 2 cores 4 cores SPARC T1 AMD Barcelona Cell BE Intel Prototype 8 cores 4 cores 8 + 1 cores 80 cores 27

Course Staff • Instructor: Steven Swanson • Lectures Tues + Thurs • TA: Hung-Wei Tseng • See the course web page for contact information and office hours. 28

Who am I? • BA/BS at University of Puget Sound • PhD at the University of Washington • Computer architecture • Ubiquitous computing • Thesis: “The WaveScalar Architecture” • At UCSD since 2006 • Heterogeneous architectures • Non-volatile, solid-state memories • Multi-processor memory system optimizations