In 1995, the top-grossing film in the U.S. was Batman Forever. (Val Kilmer as Batman, Jim Carrey as the Riddler, Tommy Lee Jones as Two-Face. Yeah.) The L.A. Rams were moving back to St. Louis, and Michael Jordan was moving back to the Bulls. Violence was rife in the Balkans. The O.J. trial happened.
It was a very different time, to be sure. But all that was nothing compared to how different the world of supercomputing was.
The Top500 list from June 1995 shows just how far the possibilities of silicon have come in the past 20 years. Performance figures are listed in gigaflops, rather than the teraflops of today, meaning that, for example, the 10th-place entrant in this week’s newly released list is more than 84,513 times faster than its two-decades-ago equivalent.
Here’s a quick run-down of the respective top 10 lists:
#10: 1995 – Cray T3D-MC512-8, Pittsburgh Supercomputing Center, 50.8 GFLOP/S
2015 – Vulcan, Lawrence Livermore National Laboratory, 4,293,300 GFLOP/S
The Pittsburgh Supercomputing Center is still an active facility, though none of its three named systems – Sherlock, Blacklight and Anton – appear on the latest Top500 list. The last time it was there was 2006, with a machine dubbed Big Ben placing 256th. (The PSC’s AlphaServer SC45 took second place in 2001 with a speed of 7,266 gigaflops.)
#9: 1995 – Cray T3D-MC512-8, Los Alamos National Laboratory, 50.8 GFLOP/S
2015 – JUQUEEN, Forschungzentrum Juelich, 5,008,900 GFLOP/S
Yes, it’s the same machine twice, which demonstrates that supercomputers were less likely to be bespoke systems filling giant rooms of their own, and more likely to be something you just bought from Cray or Intel. JUQUEEN is more than 98,600 times as powerful as the old T3D-MC512-8, a 512-core device that appears to have been more or less contained to a pair of big cabinets.
#8: 1995 – Thinking Machines CM-5/896, Minnesota Supercomputer Center, 52.3 GFLOP/S
2015 – Stampede, Texas Advanced Computing Center, 5,168,100 GFLOP/S
Thinking Machines was an early supercomputer manufacturer, based in the Boston area, that had actually gone bankrupt already by the time the June 1995 Top500 list was published – Sun Microsystems would eventually acquire most of its assets in a 1996 buyout deal. The University of Minnesota’s HPC department is now the Minnesota Supercomputing Institute, whose new Mesabi system placed 141st on the latest list at 4.74 teraflops.
#7: 1995 – Fujitsu VPP500/42, Japan Atomic Energy Research Institute, 54.5 GFLOP/S
2015 – Shaheen II, King Abdullah University of Science and Technology, 5,537,000 GFLOP/S
Fujitsu’s been a fixture on the Top500 since the list was first published in 1993, and 1995 was no exception, with the company picking up three of the top 10 spots. The Japan Atomic Energy Research Institute has dropped off the list since 2008, though it may be set to return soon, with the recent announcement that it had agreed to purchase a Silicon Graphics ICE X system with a theoretical top speed of 2.4 petaflops – which would place it just outside the top 25 on the latest list.
#6: 1995 – Thinking Machines CM-5/1056, Los Alamos National Laboratory, 59.7 GFLOP/S
2015 – Piz Daint, Swiss National Supercomputing Center, 6,217,000 GFLOP/S
For the record, we’re well over the 100,000x performance disparity between these two systems at this point. One thing that’s notable about 1995’s systems compared to today’s is the small number of cores – the CM-5 that placed sixth in 1995 used 1,056 cores, and the Fujitsu behind it used only 42. Per-core performance is still orders of magnitude higher today, but it’s worth noting that a huge proportion of the total performance increase is due to the vastly higher number of processor cores in use – no system on the 2015 list had fewer than 189,792, counting accelerators.
#5: 1995 – Fujitsu VPP500/80, Japan National Laboratory for High Energy Physics, 98.9 GFLOP/S
2015 – Mira, Argonne National Laboratory, 8,586,600 GFLOP/S
The power factor is back down to about 87,000 with the substantial jump in performance up to the 80-core Fujitsu’s nearly 100 gigaflop mark. The VPP500/80 would remain on the list through 1999, never dropping below the 90th position.
#4: 1995 – Cray T3D MC1024-8, undisclosed U.S. government facility, 100.5 GFLOP/S
2015 – Fujitsu K Computer, RIKEN Advanced Institute for Computational Science, 10,510,000 GFLOP/S
The T3D MC1024-8 system used at an undisclosed government facility (which is almost certainly not the NSA, of course) was the first on the 1995 list to top the 100 gigaflop mark, and stayed on the Top500 until 2001. That’s a solid run, and one that the Fujitsu K computer, on its fourth year in the top 5, could do well to emulate.
#3: 1995 – Intel XP/S-MP 150, Oak Ridge National Laboratory, 127.1 GFLOP/S
2015 – Sequoia, Lawrence Livermore National Laboratory, 17,173,200 GFLOP/S
The Department of Energy’s strong presence on the upper rungs of the Top500 list is one thing that hasn’t changed in 20 years, it seems – four of the top 10 in both 2015 and 1995 were administered by the DOE. The XP/S-MP 150 system boasts roughly three times as many processor cores than all but one other entry on the list, at 3,072, in a sign of things to come.
#2: 1995 – Intel XP/S140, Sandia National Laboratory, 143.4 GFLOP/S
2015 – Titan, Oak Ridge National Laboratory, 17,590,000 GFLOP/S
Indeed, the other Intel system on the 1995 list was the only other one with more cores, at 3,608. It’s even starting to look more like a modern supercomputer.
#1: 1995 – Fujitsu Numerical Wind Tunnel, National Aerospace Laboratory of Japan, 170 GFLOP/S
2015 – Tianhe-2, China National Supercomputer Center, 33,862,700 GFLOP/S
The Numerical Wind Tunnel, as the name suggests, was used for fluid dynamics simulations in aerospace research, most notably the classic wind tunnel testing to measure stability and various forces acting on an airframe at speed. The 2015 winner, China’s Tianhe-2, is almost two hundred thousand times as powerful, however.