Edit: Are they still the fastest in the world?
Originally posted January 29, 2019 on AIXchange
We all know that Summit and Sierra are the world’s fastest supercomputers, and that they run on POWER9 processors connected to NVIDIA GPUs. (The second half of this post goes into detail.)
Here’s more from CNet:
The US now can claim the top two machines on a list of the 500 fastest supercomputers, as Sierra, an IBM machine for nuclear weapons research at Lawrence Livermore National Laboratory, edged out a Chinese system that last year was the very fastest.
The Top500 list ranks supercomputers based on how quickly they perform a mathematical calculation test called Linpack. The top machine, IBM’s Summit at Oak Ridge National Laboratory, had claimed the No. 1 spot in June with a speed of 122.3 quintillion mathematical operations per second, or 122.3 petaflops.
But an upgrade gave it a score of 143.5 petaflops on the newest list. To match that speed, each person on the planet would have to perform 19 million calculations per second. Sierra got an upgrade, too, boosting its performance from 71.6 petaflops to 94.6 petaflops and lifting it from third place to second.
Summit and Sierra are siblings, each using IBM POWER9 processors boosted by Nvidia Tesla V100 accelerator chips and connected with Mellanox high-speed Infiniband network connections. They’re gargantuan machines made of row after row of refrigerator-size computing cabinets. Summit has 2.4 million processor cores and Sierra has 1.6 million.
Supercomputers are used for tasks like virtual testing of nuclear weapons, aerodynamic modeling of aircraft, understanding the formation of the universe, researching cancer and forecasting climate change effects. They’re expensive but prestigious machines that can keep scientists and engineers at the vanguard of research.
But there’s still more to this story, and, not surprisingly, it speaks to the singular quality of IBM Power Systems hardware. Top500, the supercomputer ratings group referenced in the article, published some detailed data, which I’ve condensed to this simple table below. (Go here to see the original.)
| Rank | System Cores | Rmax (TFlop/s) | Rpeak (TFlop/s) | Power (kW) |
| 1 | 2,397,824 | 143,500.0 | 200,794.9 | 9,783 |
| 2 | 1,572,480 | 94,640.0 | 125,712.0 | 7,438 |
| 3 | 10,649,600 | 93,014.6 | 125,435.9 | 15,371 |
| 4 | 4,981,760 | 61,444.5 | 100,678.7 | 18,482 |
Note the stark differences in the number of system cores deployed, as well as the power consumption. The Power Systems machines require far fewer cores and consume roughly 50 percent as much energy.
Less than half the cores, nearly half the power, and better results? These are some spectacular numbers. If you ever need to make an argument for Power Systems hardware running AIX, IBM i or Linux with the latest processors, this is tremendous ammunition.