University of California Davis - PDF document

1 Matthew Farrens University of Califor
Matthew Farrens University of California, Davis WETI Cross - Cutting Tools 2 • Began by looking at optical for off - chip communication – Logical place to start • Became intrigued by on - chip possibilities for optical – Multicores will

2 require high bandwidth interconnect â€
require high bandwidth interconnect – Electrical unlikely to scale well • Latency, power consumption issues Origins 3 • Optical networks have many upsides – High switching speeds – Energy efficiencies in 10’s of fJ /bit • Also, u

3 nique capabilities – Waveguides can
nique capabilities – Waveguides can intersect – High signal density (128 / waveguide, for example) Optics 4 • Many interesting papers in literature • Focus on basic building blocks – Waveguides, resonators, etc • Proposed designs us

4 ed large number of resonators – Hund
ed large number of resonators – Hundreds to millions – Treated as black boxes • Primarily implemented existing ideas using Optics Optical On - chip Networks 5 • Thermals important when devices sensitive to heat – Resonators drift .09nm

5 /C • Pumping many watts of power int
/C • Pumping many watts of power into chip • Is proposed network actually even thermally stable? Nothing on Thermals 6 • No tools existed to answer that question – So we built one • Well, Chris Nitta built one, anyway. Thermally Stable?

6 7 • Tool to model network power c
7 • Tool to model network power consumption • Based on Orion, CACTI – Updated and added optical modeling • Calls Hotspot to calculate temperatures • Written to be called as library like Hotspot Mintaka 8 • Validated to the best o

7 f our ability • Nothing to compare i
f our ability • Nothing to compare it to directly • Checked what we could – Compared Electrical calculations to CACTI – Compared optical numbers to published values Mintaka 9 • Model existing systems in much greater detail • Discover

8 unexpected behavior – Trimming not f
unexpected behavior – Trimming not fixed cost – Power can be absent when needed with photonic tokens • Identify interesting new challenges – Last Mile • Explore unique capabilities of Optical Interconnects – Directly Connected Optical Fab

9 ric Using Mintaka We Could 16x16
ric Using Mintaka We Could 16x16 Arbitration - Free Directly Connected Optical Fabric C. Nitta, M. Farrens , V Akella DCAF – Directly Connected Arbitration Free Photonic Crossbar for Energy - Efficient High Performance Computing , IPDPS, Shang

10 hai, May 2012 (To appear) 11 • Low
hai, May 2012 (To appear) 11 • Low - level simulations important to provide parameters • System - level modeling very important as well – Modeling small number of devices not enough Lessons Learned 12 • Need for universally agreed - upon se

11 t of values • Ideally, 3 numbers for
t of values • Ideally, 3 numbers for each parameter – Theoretical Min/Max – Experimentally demonstrated – Agreed - up/consensus number Lessons Learned 13 • Importance of thermal modeling – Particularly for technologies that are sensit

12 ive to thermals – Even electronics a
ive to thermals – Even electronics are now sensitive because of leakage Lessons Learned 14 • Better thermal modeling of ambient temperature – Need better models of amount of heat that can be moved – Thermals should be integrated with po

13 wer calculations • not separate call
wer calculations • not separate call • Need equalized /electrical interconnect models – How to accurately estimate electrical interconnect cost? – No available models for equalized electrical interconnects • Equalized on - chip interconnect t

14 oolbox Things we need 15 • Need
oolbox Things we need 15 • Need tools to model resilience – Newer, immature technologies prone to more errors • Need to model failures for new technology – Will require techniques to provide reliability • Need ways to model the effective

15 ness of techniques • Need a “Wikiâ
ness of techniques • Need a “Wiki” • Where community can contribute and create universal parameters Things we need 16 • Not looking at entire system can lead to problems – Thermal modeling often not done because it’s complicated •

16 Ignore thermal modeling at your own peri
Ignore thermal modeling at your own peril! • Not including adequate information leads to problems – Traces need dependency information, for example • Need to look at interfaces between technologies also – Last mile problem • GIGO Conclusi