Tom Kelliher, CS 220
Aug. 29, 2011
Read 1.4.
Performance measurement.
What is computer organization and why is it important? The three dimensions involved in optimizing traditional performance:
A new performance criterion: power. Determined by voltage, transistor count, clock rate.
Examples of problems to be solved:
Questions to consider:
Moore's law: the number of transistors on a chip doubles every two years. What has this given us?
Some more recent figures:
Processor | Year | Transistor Count |
AMD Athlon 64 | 2003 | 105,900,000 |
Intel Core 2 Duo | 2006 | 291,000,000 |
Intel Core 2 Quad | 2006 | 582,000,000 |
NVIDIA G80 | 2006 | 681,000,000 |
Intel Dual Core Itanium 2 | 2006 | 1,700,000,000 |
Six Core Xeon 7400 | 2008 | 1,900,000,000 |
AMD RV770 | 2008 | 956,000,000 |
NVIDIA GT200 | 2008 | 1,400,000,000 |
Eight Core Xeon Nehalem-EX | 2010 | 2,300,000,000 |
10 Core Xeon Westmere-EX | 2011 | 2,600,000,000 |
AMD Cayman | 2010 | 2,640,000,000 |
NVIDIA GF100 | 2010 | 3,000,000,000 |
Altera Stratix V | 2011 | 3,800,000,000 |
(G80: 128 stream processors -- FPUs; RV770 800 SPs; GT200 240 SPs; GF100 512 SPs)
What have architects done with these transistors?
CPUs: lots of transistors tied up in caches.
GPUs: FPU-intensive.
Interestingly, it's not easy to get transistor counts for mobile platforms, such as the OMAP 4. Why?
Most CPU sales: ARM processors, in cell phone handsets
Layered system design:
Compilation process:
One-to-one correspondence to machine code (usually).
Components of a computer:
Hierarchy:
Technologies:
Memory and I/O can't always keep up.
Buses:
Relative performance of various technologies: