China Extends Lead in Number of TOP500 Supercomputers, US Holds on to Performance Advantage

BERKELEY, Calif.; FRANKFURT, Ger­ma­ny; and KNOXVILLE, Tenn.— The 54th edi­ti­on of the TOP500 saw Chi­na and the US main­tai­ning their domi­nan­ce of the list, albeit in dif­fe­rent cate­go­ries. Mean­while, the aggre­ga­te per­for­mance of the 500 sys­tems, based on the High Per­for­mance Lin­pack (HPL) bench­mark, con­ti­nues to rise and now sits at 1.65 exa­flops. The ent­ry level to the list has risen to 1.14 peta­flops, up from 1.02 peta­flops in the pre­vious list in June 2019.

Top-ran­ked systems

The top of the list remains lar­ge­ly unch­an­ged. In fact, the top 10 sys­tems are unch­an­ged from the pre­vious list.

Sum­mit and Sier­ra remain in the top two spots. Both are IBM-built super­com­pu­ters employ­ing Power9 CPUs and NVIDIA Tes­la V100 GPUs. Oak Ridge Natio­nal Laboratory’s Sum­mit sys­tem holds top honors with an HPL result of 148.6 peta­flops. The second-ran­ked Sier­ra sys­tem at Law­rence Liver­mo­re Natio­nal Labo­ra­to­ry comes in at 94.6 petaflops.

Clo­se behind at num­ber three is the Sun­way Tai­hu­Light super­com­pu­ter, with an HPL mark of 93.0 peta­flops. Tai­hu­Light was deve­lo­ped by China’s Natio­nal Rese­arch Cen­ter of Par­al­lel Com­pu­ter Engi­nee­ring & Tech­no­lo­gy (NRCPC) and is instal­led at the Natio­nal Super­com­pu­ting Cen­ter in Wuxi. It is powered exclu­si­ve­ly by Sunway’s SW26010 processors.

Tian­he-2A (Mil­ky Way-2A), a sys­tem deve­lo­ped by China’s Natio­nal Uni­ver­si­ty of Defen­se Tech­no­lo­gy (NUDT) and deploy­ed at the Natio­nal Super­com­pu­ter Cen­ter in Guang­zhou, Chi­na, holds the num­ber four spot with 61.4 peta­flops. It is powered by Intel Xeon CPUs and Matrix-2000 accelerators.

Fron­te­ra, a Dell C6420 sys­tem, keeps it num­ber five ran­king, with an HPL result of 23.5 peta­flops. Powered sole­ly by Xeon Pla­ti­num pro­ces­sors, it was instal­led at the Texas Advan­ced Com­pu­ting Cen­ter of the Uni­ver­si­ty of Texas in 2018.

At num­ber six is Piz Daint, a Cray XC50 super­com­pu­ter instal­led at the Swiss Natio­nal Super­com­pu­ting Cent­re (CSCS) in Luga­no, Switz­er­land. At 21.2 peta­flops, it remains the most powerful sys­tem in Europe.

Tri­ni­ty, a Cray XC40 sys­tem ope­ra­ted by Los Ala­mos Natio­nal Labo­ra­to­ry and San­dia Natio­nal Labo­ra­to­ries is ran­ked num­ber seven, with an HPL per­for­mance of 20.2 peta­flops.  It is powered by Intel Xeon and Xeon Phi processors.

Lis­ted at num­ber eight is the AI Bridging Cloud Infra­struc­tu­re (ABCI), instal­led at Japan’s Natio­nal Insti­tu­te of Advan­ced Indus­tri­al Sci­ence and Tech­no­lo­gy (AIST). The Fuji­tsu-built sys­tem is equip­ped with Intel Xeon Gold pro­ces­sors and NVIDIA Tes­la V100 GPUs, achie­ving an HPL result of 19.9 petaflops.

SuperMUC-NG is in the num­ber nine posi­ti­on with 19.5 peta­flops. It’s instal­led at the Leib­niz-Rechen­zen­trum (Leib­niz Super­com­pu­ting Cent­re) in Gar­ching, near Munich. The Leno­vo-built machi­ne is powered by Intel Pla­ti­num Xeon processors.

The Las­sen super­com­pu­ter is ran­ked num­ber 10, deli­ve­ring 18.2 peta­flops. Instal­led at Law­rence Liver­mo­re Natio­nal Labo­ra­to­ry, Las­sen is the unclas­si­fied coun­ter­part to the clas­si­fied Sier­ra sys­tem and shares the same IBM Power9/NVIDIA V100 GPU architecture.

The most powerful new super­com­pu­ter on the list is AiMOS, which shows up at posi­ti­on 25, tur­ning in an HPL result of 8.0 peta­flops. The IBM-built sys­tem is instal­led at the Rens­se­laer Poly­tech­nic Insti­tu­te Cen­ter for Com­pu­ta­tio­nal Inno­va­tions (CCI), and like Sum­mit, Sier­ra, and Las­sen, it is equip­ped with Power9 CPUs and NVIDIA V100 GPUs. 

Geo­gra­phi­cal breakdown

The num­ber of TOP500 instal­la­ti­ons in Chi­na con­ti­nues to rise and now sits at 227, up from 219 six months ago. Mean­while, the share of US-based sys­tem remains near its all-time low at 118. Howe­ver, sys­tems in the US are on avera­ge signi­fi­cant­ly lar­ger, which trans­la­ted to a 37.8 per­cent share of the list’s aggre­ga­te per­for­mance. Chi­na is clo­se behind with a 31.9 per­cent per­for­mance share. Howe­ver, com­pared to six months ago, this per­for­mance gap has shrunk. The June 2019 list had the US with a 38.4 per­cent of the list’s aggre­ga­te per­for­mance and Chi­na with 29.9 percent.

Japan remains in third place in the num­ber of TOP500 sys­tems, with 29, fol­lo­wed by France with 18, Ger­ma­ny with 16, the Net­her­lands with 15, Ire­land with 14, and the United King­dom with 11. All other count­ries were in the sin­gle digits.

Ven­dor shares

 As a reflec­tion of China’s domi­nan­ce in sheer num­bers, the top three sys­tem ven­dors with regard to the num­ber of instal­la­ti­ons are Leno­vo, (174), Sugon (71), and Inspur (65). Cray is num­ber four, with 36 sys­tems, and HPE is num­ber five, with 35. Note that Cray is now part of HPE, so taken tog­e­ther they would effec­tively tie Sugon with 71 systems.

At the chip level, Intel con­ti­nues its domi­nan­ce. Its pro­ces­sors are pre­sent in 470 of the 500 sys­tems, split bet­ween mul­ti­ple gene­ra­ti­ons of Xeon and Xeon Phi hard­ware. IBM is second with 14 sys­tems – 10 with Power CPUs and four with Blue Gene/PowerPC CPUs. AMD claims just three sys­tems on the cur­rent list.

The­re are now two Arm-based super­com­pu­ters on the list: the Astra sys­tem deploy­ed at San­dia Natio­nal Labo­ra­to­ries, which is equip­ped with Marvell’s ThunderX2 pro­ces­sors, and Fujitsu’s A64FX pro­to­ty­pe sys­tem, a pre­cur­sor to the Fug­a­ku (Post‑K) exas­ca­le sys­tem desti­ned for RIKEN in 2021. The Fuji­tsu machi­ne is new to the list and occu­p­ies posi­ti­on 159, with an HPL per­for­mance of 2.0 petaflops.

NVIDIA is the domi­nant ven­dor for acce­le­ra­tors. Its GPUs are pre­sent in 136 of the 145 acce­le­ra­ted sys­tems. On the pre­vious list six months ago, the­re were 134 acce­le­ra­ted systems.

Inter­con­nect breakdown

Ether­net is used in 52 per­cent (258) of the TOP500 sys­tems, while Infi­ni­Band is the net­work-of-choice in 28 per­cent (140) of sys­tems. Howe­ver, from a per­for­mance per­spec­ti­ve, tho­se posi­ti­ons are rever­sed, with Infi­ni­Band-based machi­nes repre­sen­ting 40 per­cent of the TOP500’s aggre­ga­te per­for­mance and Ether­net-based machi­nes with 29 per­cent. Cus­tom inter­con­nects, with just 46 instal­la­ti­ons, cla­im 22 per­cent of the list’s instal­led performance.

Green500 results

The rela­ted Green500 list, which mea­su­res ener­gy-effi­ci­en­cy of the world’s top super­com­pu­ters, has chan­ged con­sider­a­b­ly com­pared to six months ago. Num­ber one on the list is the afo­re­men­tio­ned A64FX pro­to­ty­pe super­com­pu­ter, which deli­ver­ed 16.9 gigaflops/watt. Just slight­ly less green at num­ber two sys­tem is NA‑1, a Zet­tas­ca­ler machi­ne that uses PEZY Computing’s PEZY-SC2 pro­ces­sors and deli­vers 16.3 gigaflops/watt. It is being rea­di­ed for a future instal­la­ti­on at NA Simu­la­ti­on in Japan.

The third-ran­ked Green500 sys­tem is the new AiMOS sys­tem from IBM, fol­lo­wed by two more IBM sys­tems based on the same Power9/NVIDIA V100 design: Sato­ri at 15.6 gigaflops/watt and Sum­mit at 14.7 gigaflops/watt. The remai­ning top 10 Green500 sys­tems on the list – AI Bridging Cloud Infra­struc­tu­re, MareN­os­trum P9 CTE, TSUBAME 3.0, PANGEA III, and Sier­ra – also use NVIDIA GPU accelerators.

HPCG results

The two top-ran­ked Sum­mit and Sier­ra super­com­pu­ters on the TOP500, also remain in the top two spots on the list based on the High-Per­for­mance Con­ju­ga­te Gra­di­ent (HPCG) bench­mark. Sum­mit achie­ved 2.93 HPCG-peta­flops, with Sier­ra at 1.80 HPCG-peta­flops. All the remai­ning top 10 HPCG ent­ries, deli­ver­ed less than one HPCG-peta­flops. With the excep­ti­on of the now-decom­mis­sio­ned K‑computer, all 10 of the­se sys­tems car­ri­ed over from the pre­vious list six months ago.

About the TOP500 List

The first ver­si­on of what beca­me today’s TOP500 list star­ted as an exer­cise for a small con­fe­rence in Ger­ma­ny in June 1993. Out of curio­si­ty, the aut­hors deci­ded to revi­sit the list in Novem­ber 1993 to see how things had chan­ged. About that time, they rea­li­zed they might be onto some­thing and deci­ded to con­ti­nue com­pi­ling the list, which is now a much-anti­ci­pa­ted, much-wat­ched, and much-deba­ted twice-year­ly event.