For the first time since the TOP500 group began publishing their list of the fastest computers in the world, there was no turnover in the top 10 machines. In fact, the only change at the top was the new record Linpack mark set by the now fully deployed K Computer at RIKEN. After installing the remaining racks over the last six months, Fujitsu expanded that system’s performance from 8.16 to 10.51 petaflops, a value that exceeds the next seven fastest systems combined.
The current top of the list looks like this:
- 10.51 petaflops, K computer, Japan
- 2.57 petaflops, Tianhe-1A, China
- 1.76 petaflops, Jaguar, United States
- 1.27 petaflops, Nebulae, China
- 1.19 petaflops, TSUBAME 2.0, Japan
- 1.11 petaflops, Cielo, United States
- 1.08 petaflops, Pleiades, United States
- 1.05 petaflops, Hopper, United States
- 1.05 petaflops, Tera-100, France
- 1.04 petaflops, Roadrunner, United States
But despite the inactivity at the top, aggregate performance for the whole list ticked up quite respectably from 58.7 petaflop to 74.2 petaflops over the last six months. And, as is usually the case, the bottom of the list experienced lots of turnover. The new entry point into the list is now 50.94 teraflops, more than 10 teraflops more than was required in June. In fact the new number 500 system, an HP Proliant cluster for a “IT Service Provider,” was sitting at number 301 just six months ago.
From a geographic point of view, the US still owns the majority of the 500 fastest supercomputers, with 263, adding 8 more systems since June. In second place is China with 74, having added 13 new machines over the last six months. Japan (30), the UK (27), France (23) Germany (20), Canada (9), Poland (6), Russia (6), and Australia (4) round out the top 10.
GPUs are continuing to gain share at a health clip, with a total of 37 systems now sporting the graphics chips — about twice as many as there were in June. NVIDIA parts are in 35 of these systems, while AMD (ATI) GPUs managed just two appearances.
But the x86 CPU is still the king of HPC. Intel Xeons processors have the lion’s share of the CPUs on the tops supers, appearing in 76.8 percent of the systems. The majority of those are Westmere-EP (Xeon 5600 series) processors. AMD Opterons represent just 12.6 percent of the total, while IBM Power chips, of one sort or another, are in 9.2 percent of the machines.
Perhaps the most discouraging trend is that of power consumption at the top of the list. Aggregate wattage for the fastest 10 machines is 4.56 MW, up slightly from 4.3 MW in June. More importantly, the average power efficiency of the elite supercomputers is 464 megaflops/watt, the same as it was six months ago.
For those looking forward to exascale supercomputers before the end of this decade, that’s rather disheartening. There is a general consensus that exaflop machines should consume no more than 20 MW, which translates into 50 gigaflops/watt. Not only are current power efficiencies two orders of magnitude off, but for the time being at least, progress in this area seems to have stopped. As plans for exascale computers start to solidify in the next few years, look for this TOP500 metric to come under increasing scrutiny.