With the kick-off of the International Supercomputing Conference (ISC) in Frankfurt this morning, the 53rd Top500 list made its debut, and this one’s for petafloppers only. The entry point for the new list is 1.022 petaflops, and the total aggregate performance of all 500 systems for the first time exceeds 1.5 exaflops. The top new entry to the list is Frontera, the NSF-funded Dell EMC system installed at the Texas Advanced Computing Center (TACC), at number five with 23.5 Linpack petaflops.

China claims the most systems (219) with 208 of these pretty evenly split among Lenovo (74), Inspur (71) and Sugon (63). Lenovo, which acquired IBM’s x86 business in 2014, maintains its lead as the number one Top500 vendor by number of systems (173). As we continue to note, the majority of these are not HPC systems or supercomputers; in fact, over half the full list (about 55 percent) is now comprised of web-scale/cloud systems.

The US (with 116 systems) trails China by nearly half in terms of system share, but when it comes to aggregate performance share, the US leads with 38.4 percent of list flops compared with China’s 29.9 percent. Summit and Sierra — still at number one and two respectively — enjoy 243.2 Linpack petaflops between them, accounting for nearly a sixth of the list’s aggregate capacity. Oak Ridge National Laboratory has optimized Summit’s Linpack performance, such that it now delivers 148.6 petaflops on the HPL benchmark, versus 143.5 petaflops six months ago.

The prime manufacturer behind Summit and Sierra, IBM, is in first place by aggregate list performance with 321 petaflops, ahead of Lenovo, which touts 302 total petaflops. Cray has the next highest aggregate list performance share (183.4 petaflops), followed by HPE (120.1 petaflops) and Sugon (96 petaflops). In anticipation of HPE purchasing Cray, we note that a combined union on this list would deliver a total of 303.5 aggregate petaflops — still shy of IBM’s total, and a smidgen ahead of Lenovo’s.

Restricting our analysis to the top 100 systems to better gauge the HPC/supercomputing picture (an imperfect but quick reckoning), the top five countries by system share are US with 37, Japan with 15, China and Germany tied with 9, followed by France with six systems (including two Jean Zay systems, an accelerated machine in 42^nd position and a CPU-only one in 72^nd). By aggregate performance, US leads the top 100 pack with 49 percent of the flops, China with 17 percent, Japan with 10 percent, and Germany and France both with 5 percent. The top vendors by system share in the top 100 grouping are Cray with 25, HPE with 18, IBM with eight [Ed. this is the corrected number], Fujitsu with seven, and then a three-way tie between Dell EMC, Nvidia and Sugon (all with three systems). By aggregate performance share, the order changes to IBM, Cray, HPE, Fujitsu and Dell EMC.

View from the Top: Bienvenida Frontera

Frontera, the Dell C6420 system installed at the TACC, makes its list debut at number five, achieving 23.5 Linpack petaflops (38.7 peak petaflops) using 448,448 Intel Xeon cores. Frontera’s 8,000 nodes are hooked together using Mellanox HDR InfiniBand. This is the debut list for HDR, which connects three other systems (#68, #166, and #264) in addition to Frontera.

Lassen — Sierra’s unclassified “companion” system at LLNL that was built by IBM, Nvidia and Mellanox — moves up into the number 10 spot with an upgraded capacity of 18.2 Linpack petaflops.

Titan, ORNL’s Cray XK7 system, which had a long reign as the United States’ fastest supercomputer since its debut as an upgrade to Jaguar in November 2012 until Summit was deployed in June 2018, has shifted out of the top 10, as has Sequoia, the last BlueGene/Q system at LLNL.

Taking Lassen’s previous spot in 11^th place is another IBM system, a new entrant to the list and the fastest commercial supercomputer: Pangea III. Measured at 17.9 Linpack petaflops (rPeak 25 petaflops), the system was built by IBM for Total to help the energy company more accurately locate new oil reservoirs and better assess the potential associated revenue opportunities.

One vendor missing from the top 10 pack is HPE, an observation made by market analyst Addison Snell, CEO of Intersect360 Research. “IBM, Cray, Fujitsu, Lenovo, and Dell EMC all have top ten systems. The biggest HPE system is no slouch, at number 17, the HPC4 system at Italian energy company Eni. And of course, HPE will soon acquire Cray, inheriting the plans for the first two exascale systems in the US,” Snell commented to HPCwire.

How the Chips Fall

While there are more CPU options appearing in the general server market, the trend has not had much of an impact on the Top500 yet. Intel chips still dominate in so far as they appear in 95.6 percent of all systems, although in some cases, much or most of the heavy lifting is owed to accelerators or coprocessors. IBM Power processors are in seven systems, followed by AMD CPUs, which are used in three systems. A single supercomputer on the list, Astra at Sandia National Labs, is powered by (Marvell/Cavium) Arm processors. Built by HPE, Astra debuted on the November 2018 list at 204 with 1.5 petaflops. An upgrade goosed Linpack to 1.76 petaflops, elevating the system into 156^th position.

Accelerated Computing

There are 133 systems that use accelerator or co-processor technology, dipping slightly from 138 six months ago. Overall the trend toward acceleration remains strong with Nvidia seeing 25 percent growth in the number of Top500 systems that employ its GPUs in the last year. The total number of Nvidia GPU-accelerated systems has gone from 98 last June 2018 to 125 total systems on the current list.

With AMD and Intel server GPUs coming into the market, or preparing to in Intel’s case — including already announced big wins (Aurora, Frontier) — this will be an interesting trend to follow.

Nvidia has two DGX-2 systems on the list, Circe and DGX SuperPod, which integrate Xeon Skylake Platinum 8174 processors, Nvidia V100 GPUs, NVLink technology and Mellanox EDR InfiniBand. Both are used internally for research and company workloads.

Circe, a cluster of 36 DGX-2H boxes that ranks 71 with 3.1 Linpack petaflops, entered the list last November. DGX SuperPod, comprising 96 DGX-2H nodes, joins the list today at number 22 with a Linpack score of 9.4 petaflops. Nvidia used an earlier version of this system for its MLPerf benchmarking.

“This is our system that we use to develop most notably the brains for our autonomous vehicle platform,” said Ian Buck, general manager and vice president of accelerated computing at Nvidia, referring to the DGX SuperPod. He noted that the 96 DGX systems are highly connected. “We have more than 10 Mellanox EDR InfiniBand NICs per node connected together with multiple direction level switches, connecting more than 1,500 tensor core GPUs.”

Hyperscale Takeover Continues

A tipping point has been reached, where the number of IT/web-scale machines on the Top500 now outnumber true HPC clusters and supercomputers. Claims of list-stuffing were raised last year at ISC due to an increase in hyperscale systems being creatively benchmarked for the list (which to be clear is not breaking any rules and is not a new practice; it dates back to at least 2010). The renewed enthusiasm for running the HPL benchmark on cloud systems that will be deployed and leveraged in a disaggregrated manner is partly behind China’s ascendance on the list as Chinese vendors Sugon, Inspur and Lenovo have made a concerted effort to get their systems on the list. Somebody has to lead these efforts; they don’t just happen.

In-line with this trend, Gigabit Ethernet (GE) continues to expand its hold on the list, connecting 271 systems (a 54.2 share). By aggregate flops, GE carries just 25 percent of list share. The other 75 percent of flops were conducted via InfiniBand, OmniPath, custom and proprietary interconnects.

New Entrant Profile

There are 94 new entrants to the list, bringing in 223 petaflops with China backing 35 percent of these flops and US responsible for 30 percent. Of the 94 additions, 45 are Chinese and 17 are US. Roughly two-thirds of the new systems are cloud/web-scale machines.

Green500 & HPCG

The Green500, which tracks the most energy-efficient Top500 machines, has been integrated into the Top500 reporting process and the High Performance Conjugate Gradients (HPCG) benchmark is also included in the list.

Maintaining their positions from six months ago, the top two Green500 systems are Shoubu system B and DGX SaturnV Volta ranked 471 and 469 on the Top500, respectively. Shoubu system B, made by PEZY/Exascalar and located at RIKEN, achieves 17.6 gigaflops/watt; while DGX SaturnV Volta, made by Nvidia for Nvidia, delivers 15.1 gigaflops/watt.

Top500 champ Summit, located at Oak Ridge National Laboratory in Tennessee, is in third place on the Green500 with a slightly improved result of 14.71 gigaflops/watt.

At number four is the AI Bridging Cloud Infrastructure (ABCI) system installed at the National Institute of Advanced Industrial Science and Technology (AIST) in Japan. Number seven on the Top500, ABCI delivers 14.4 gigaflops/watt.

The MareNostrum P9 CTE, moves up three spots into fifth place with a refreshed result of 14.1 gigaflops/watt. Built by IBM for Barcelona Supercomputing Center, the machine delivers 1.14 Linpack petaflops for a 393^rd ranking on the Top500.

TSUBAME 3.0 drops from fifth to sixth place on the Green500 with a result of 13.7 gigaflops/watt. The 8.1 petaflops supercomputer, 25^th on the current Top500 list, was made by HPE for GSIC Center, Tokyo Institute of Technology.

Total’s new supercomputer, Pangea III, new to the Top500 list in 11^th place, nabs the seventh position on the Green500, achieving 13.07 gigaflops/watt.

Sierra, LLNL’s flagship system that is number two on the Top500, moves to eighth place, maintaining its efficiency score of 12.72 gigaflops/watt.

Sugon’s Advanced Computing System (PreE) and Taiwania 2 (installed at Taiwan’s National Center for High Performance Computing) captured the nine and ten positions on the Green500, respectively. The PreE system—number 43 on the Top500 with 4.3 petaflops—is powered by the Dhyana CPU, a Hygon-licensed implementation of AMD’s Epyc processor. Built by Quanta Computer, Taiwania 2 occupies spot number 23 on the Top500 with a Linpack rating of 9 petaflops.

All of the top 10 greenest supercomputers employ accelerators (primarily Nvidia GPUs) in a hybrid system design. In fact, the highest-ranking Green500 CPU-only machine is Sunway TaihuLight in 25th place. China’s TaihuLight, the pride of Wuxi, home to the computer’s eponymous Taihu Lake, maintains its third-spot ranking on the Top500 with 93 Linpack petaflops.

One of the encouraging things about the Green500 is the diversity of international representation. Japan, the United States, Spain, France, China, Taiwan, the United Kingdom, Italy, Canada, Switzerland and Brazil all have machines in the top 25.

On the HPCG, Summit and Sierra hold the top positions with Japan’s K computer at number three. Respectively, they achieved 2.93 HPCG-petaflops, 1.80 HPCG-petaflops, and .60 HPCG-petaflops. As the list authors note, “Summit and Sierra remain the only two systems to exceed a petaflop on the HPCG benchmark…. The average HPCG result on the current list is 213.3 teraflops, a marginal increase from 211.2 six months ago.”