1. Created by Simpo PDF Creator Pro (unregistered version)
http://www.simpopdf.com
UFOs – Unidentified Flying Objects
Ufology – is a neologism coined to describe the collective
efforts of those who study reports and associated evidence of
unidentified flying objects (UFOs).
Ufologist – A ufo investigator is called a ufologist
Ufo Sightings- Some eye witnesses to the UFOs
Roswell Incident- called the Roswell UFO crash
1947
UFO Conspiracy – Worldwide UFO cover Ups and
related theories
Alien- An extra-terrestrial being is called an alien
2. Lost in the Bermuda Triangle:
Complexity vs. Energy vs. Performance
John Paul Shen
Intel Microarchitecture Research Lab
June 18, 2006
WCED Panel
Microprocessor Performance Evolution
Frequency
Performance =
2
PathLength × CPI
John Paul Shen
John Paul Shen - June 18, 2006 2
1
3. Power/Performance (EPI) Evolution
Power = EPI × IPC × Frequency
50
Intel EPI (nj)
45
nj Microprocessors 65nm at
40 48 1.33v
Pentium 4 I =
35
(Cedarmill) EP i486 10
Power = Performance1.75 Pentium 14
30
Power
25 Pentium Pro 24
Pentium 4
20
(Willamette) Pentium 4 (WMT) 38
15
Pentium M Core Duo Pentium 4 (CDM) 48
Dothan Yonah Pentium M (Banias) 13
10 Banias
Pentium Pro
nj Merom Pentium M (Dothan) 15
5 EPI = 10
i486 Pentium Core Duo (Yonah) 11
0
0 2 4 6 8 10 Core Duo (Merom) 10
Scalar Performance
Power: single core power (relative to i486 baseline)
Performance: SPECint performance (relative to i486 baseline)
EPI: average energy spent per instruction (in nano-joules)
3 John Paul Shen
John Paul Shen - June 18, 2006 3
EPI Optimization for CMP Architectures
Power/Performance Scaling
300
EPI: CPU Cores Prog. Accelerators Fixed Function Units
10nj 1nj 0.1nj 0.01nj
250
Power = EPI × IPC × Frequency = EPI × IPS EPI = 5 nj
200 (linear power scaling)
20x Performance Increase
Power (Watt)
EPI = 1 nj
150W Power Envelope (power scales at n^1.1)
150
EPI = 1 nj
(linear power scaling)
100 EPI = 0.5 nj
EPI (linear power scaling)
Throttling
50 NP/DSP/GPU EPI
IXP2800 ~1 nj
TMS320C6713 ~0.7 nj
0
GeF7800GTX ~0.6 nj
0
5
10
15
20
25
30
35
40
45
50
55
60
65
70
75
80
85
90
95
100
105
110
115
120
125
Performance (GIPS)
Performance (GIPS) Intel Gen4 ~0.3 nj
4 John Paul Shen
John Paul Shen - June 18, 2006 4
2
4. Research Challenges
• 10x Reduction of Core EPI:
– Avoid O(n2) and O(n3) structures
– Leverage heterogeneous cores/accelerators
• Linear Power Scaling of Uncore:
– Provide on-demand interconnects
– Eliminate legacy interfaces
• 2x Reduction of Design Cycle:
– Adopt modular design style
– Reuse building blocks
5 John Paul Shen
John Paul Shen - June 18, 2006 5
3