Maynard (MA) - Dr. Wu-chun Feng of Virginia Tech University created the Green500 list, a metric of performance per watt of the top 500 supercomputers in the world. By calling attention to power consumption issues, his ranking serves as a type of gage allowing outsiders to view computing power in relative terms. A company called SiCortex believes that much more is needed in this area and proposes a new Green Computing Performance Index (GCPI) that changes the way performance-per-watt is measured across an entire server farm.


Green computing

Supercomputing farms today are big power users. It is estimated that by the year 2020, worldwide server farms will consume enough energy that the greenhouse gas emissions generated by powering them will exceed the entire airline industry's emissions.

About 2/3rds of that power use goes into cooling the machines. Advances in computer technology has reduced significantly the number of processors and heat required to carry out a specific workload. Still, as advances in abilities continue so do advances in desired workloads. To gage this easily, consider the primitive graphics in movies like Tron and The Last Starfighter, compared to the graphics in today's CGI movies like WALL-E. Huge advancements made possible by greater computing power, and that means more electrical power.


New metric

SiCortex has created a blog, community forums and a website where a host of industry standard HPCC benchmarks can be run with the results being uploaded and viewed. This helps give the entire community a feel for how the GCPI operates and relates as a metric, and how various datacenters fall into place.

Seeing is believing, so here is a sample of how the new metric stacks up with some existing supercomputers. Note that the Cray Opteron-powered computer at Swiss National Supercomputing Centre CSCS is used as the baseline (left-most data column) :

	Cray	IBM	HP	SGI	Cray	SGI	SiCortex
Machine	XT3	Blue Gene	CP3000BL	Altix ICE 8200EX	XT4	Altix 8200EX	SC1458
Date	02-25 2006	04-06 2006	03-19 2008	05-09 2008	05-14 2008	06-18 2008	10-24 2008
CPUs	1100	1024	1024	512	8464	1024	1416
HPL	1.000	8.013	4.977	13.681	7.670	13.715	14.912
PTRANS	1.000	2.193	0.572	1.969	0.786	1.019	14.915
Single STREAM Copy	1.000	5.162	2.413	6.427	6.901	6.694	6.172
Single STREAM Scale	1.000	3.425	2.348	6.509	6.747	6.520	5.718
Single STREAM Add	1.000	4.280	2.333	5.453	6.711	5.634	6.154
Single STREAM Triad	1.000	4.308	2.346	5.622	6.720	5.679	6.191
EP STREAM Copy	1.000	3.583	0.459	1.706	2.435	1.707	4.861
EP STREAM Scale	1.000	3.135	0.445	1.680	2.400	1.667	4.702
EP STREAM Add	1.000	3.830	0.470	1.494	2.224	1.489	5.289
EP STREAM Triad	1.000	4.297	0.484	1.545	2.483	1.539	5.414
Single Random Access	1.000	13.842	3.769	11.559	4.208	11.530	28.547
EP Random Access	1.000	8.646	1.259	4.352	2.734	4.367	23.617
Global Random Access	1.00000	25.44604	7.97405	35.95537	7.20343	49.64298	78.61771
Random Ring 1/Latency	1.00000	43.29038	4.31850	7.81378	0.14991	4.08843	155.93199
Random Ring Bandwidth	1.00000	2.63773	0.48935	2.93413	0.12865	0.83139	8.29086
Natural Ring 1/Latency	1.00000	50.90096	9.75850	36.54946	0.42830	17.57853	164.98210
Natural Ring Bandwidth	1.00000	41.38019	11.02045	33.60891	0.48068	12.21835	228.32237
Random Ring Bandwidth	1.00000	2.63773	0.48935	2.93413	0.12865	0.83139	8.29086
1/PingPong Latency 1/Average	1.00000	45.11510	10.51735	35.83492	0.48771	13.79020	253.66355
1/PingPong Bandwidth 1/Average	1.000	3.429	3.695	13.541	0.902	5.234	84.422
1/Single DGEMM	1.000	11.448	5.146	13.894	7.965	14.059	15.431
EP DGEMM	1.000	10.870	4.866	13.408	7.830	13.574	15.054
1/Single FFT	1.000	8.066	4.508	12.229	3.838	19.768	14.545
EP FFT	0	0	0	0	0	0	0
1/Global FFT	1.000	5.334	1.273	3.341	1.548	2.390	9.984
Green Computing Performance Index	1.00	6.44	2.60	7.57	4.28	7.65	14.28

Note how well SiCortex's SC1458 fares using their new metric. Read more at the SiCortex website.