The Green500 List — November 2014

L‑CSC Emer­ges as the Most Ener­gy-Effi­ci­ent Super­com­pu­ter in the World

New Orleans, LA, USA; Novem­ber 20, 2014 – A new super­com­pu­ter, L‑CSC from the GSI Helm­holtz Cen­ter, emer­ged as the most ener­gy-effi­ci­ent (or gree­nest) super­com­pu­ter in the world, accor­ding to the 16th edi­ti­on of the twice-year­ly Green500 list of the world’s most ener­gy-effi­ci­ent super­com­pu­ters. The L‑CSC clus­ter was the first and only super­com­pu­ter on the list to sur­pass 5 gigaflops/watt (bil­li­ons of ope­ra­ti­ons per second per watt). It was powered by Intel Ivy Bridge CPUs and a FDR Infi­ni­band net­work and acce­le­ra­ted by AMD Fire­Pro™ S9150 GPUs.

In fact, the top three slots of the Green500 were powered by three dif­fe­rent acce­le­ra­tors with num­ber one, L‑CSC, being powered by AMD Fire­Pro™ S9150 GPUs; num­ber two, Sui­ren, powered by PEZY-SC many-core acce­le­ra­tors; and num­ber three, TSUBAME-KFC, powered by NVIDIA K20x GPUs. Bey­ond the­se top three, the next 20 super­com­pu­ters were also accelerator-based.

L‑CSC achie­ved the first posi­ti­on on the Novem­ber 2014 Green500 List with an impres­si­ve 5.27 giga­flops per watt. This sys­tem used Intel Ivy Bridge CPUs, AMD Fire­Pro GPUs, and ener­gy-effi­ci­ent soft­ware design to achie­ve this feat. Sui­ren, a super­com­pu­ter from the High Ener­gy Acce­le­ra­tor Rese­arch Organization/KEK in Japan, occu­p­ied the second spot at 4.95 giga­flops per watt. Like L‑CSC, it used Intel Ivy Bridge CPUs, but cou­pled them with many-core acce­le­ra­tors from PEZY-SC. Final­ly, TSUBAME-KFC, which was the most ener­gy-effi­ci­ent super­com­pu­ter over the pre­vious two edi­ti­ons of the Green500, came in third in this edi­ti­on of the Green500. TSUBAME-KFC was the first super­com­pu­ter to have bro­ken the 4 gigaflops/watt bar­ri­er. Like the other two acce­le­ra­tor-based machi­nes, it used Intel Ivy Bridge CPUs, but cou­pled with NVIDIA K20x GPUs.

Other high­lights from this Green500 List include the following:

This marks the first time that a super­com­pu­ter (L‑CSC) using AMD GPUs has held the top spot.
The Sui­ren super­com­pu­ter is the first sys­tem desi­gned with cus­tom acce­le­ra­tors to enter the top five sin­ce Novem­ber of 2010, when the GRAPE-DR super­com­pu­ter achie­ved the same feat.
The PEZY-SC many-core acce­le­ra­tors in the Sui­ren super­com­pu­ter use ARM con­trol­led cores. This repres­ents the first time that an ARM-based chip has appeared so pro­min­ent­ly on the list.

Assum­ing that L‑CSC’s ener­gy effi­ci­en­cy could be sca­led line­ar­ly to an exa­flop super­com­pu­ting sys­tem, one that can per­form one tril­li­on floa­ting-point ope­ra­ti­ons per second, such a sys­tem would con­su­me on the order of 190 mega­watts (MW). “Alt­hough this 190-mega­watt power enve­lo­pe is still far from DARPA’s opti­mi­stic tar­get of a 67-mega­watt power enve­lo­pe, it is appro­xi­m­ate­ly 16 times bet­ter than the initi­al pro­jec­tion of a near­ly 3000-mega­watt power enve­lo­pe from 2007 when the first offi­ci­al Green500 list was laun­ched,” says Wu Feng of the Green500.

About the Green500:

The Green500 has pro­vi­ded a ran­king of the most ener­gy-effi­ci­ent super­com­pu­ters in the world sin­ce Novem­ber 2007. For deca­des, the noti­on of super­com­pu­ter “per­for­mance” has been syn­ony­mous with “speed.” This par­ti­cu­lar focus has led to the emer­gence of super­com­pu­ters that con­su­me egre­gious amounts of elec­tri­cal power and pro­du­ce so much heat that extra­va­gant coo­ling faci­li­ties must be con­s­truc­ted to ensu­re pro­per ope­ra­ti­on. In addi­ti­on, when the­re is an empha­sis on speed as the ulti­ma­te metric, it often comes at the expen­se of other per­for­mance metrics, such as ener­gy effi­ci­en­cy, relia­bi­li­ty, avai­la­bi­li­ty, and usa­bi­li­ty. As a result, the­re has been an extra­or­di­na­ry increase in the total cost of owner­ship (TCO) of a super­com­pu­ter. Con­se­quent­ly, the Green500 seeks to rai­se the awa­re­ness in ener­gy effi­ci­en­cy of super­com­pu­ting and to dri­ve it as a first-order design cons­traint on par with speed.