China Tops Top500

The last time I wrote about the Green500, the Chinese machine Tianhe-2 was at the top of the Top500 list. At that point, Tianhe-2 had slipped from 49^th to 64^th on the Green500. Tianhe-2 has now slipped to second place on the Top500, only surpassed by yet another new Chinese machine, NRCPC’s Sunway TaihuLight. There has also been some consolidation between the Top500 and the Green500 lists: it was announced earlier this year on May 6 that the Top500 and Green500 merged. This is convenient for people who are tracking both of these lists, as they are now accessible from one site.

Energy Efficiency vs. Total Compute Capability

There are three very notable accomplishments about the Sunway TaihuLight:

1. Sunway TaihuLight’s Rmax score of 93 Petaflops significantly outdistances Tianhe-2’s Rmax of 34 Petaflops. The new machine is ~2.75x faster than the second place machine and ~5.28x faster than the 3^rd place Titan.
2. Sunway TaihuLight’s position on the Green list is 3^rd, only behind the two Japanese machines (Shoubu and Satsuki) made by PEZY Computing. This in itself is a remarkable achievement. Tianhe-2 has slipped to 91^st on the Green list at 1900 Mflops/Watt and Titan is 71^st at 2143 Mflops/Watt. At 6051 Mflops/Watt, Sunway TaihuLight is within 2.4% of the efficiency of the 2^nd place PEZY machine on the Green list and within almost 10% of the 1^st place PEZY machine. Sunway TaihuLight also holds nearly a 27% efficiency lead over the 4^th place machine on the Green  list (Sugon Cluster W780I) which is the most efficient x86/GPU heterogeneous machine on the Green list but only has an Rmax score of 0.31 petaflops (i.e., two orders of magnitude down from the top two Top500 machines).
3. You have to go all the way down to the 120^th position on the Top500 in order to find another machine that doesn’t use either an Intel (98), IBM (10), SPARC (7) or AMD (3) processor. The new #1 machine uses a Sunway SW26010 260C 1.45GHz designed in China. Oh, the machine at 120?  It’s the Sunway BlueLight MPP which uses the ShenWei processor SW1600 16C 975MHz. The Sunway BlueLight MPP is rated at an Rmax of 0.7 Petaflops and when it debuted, was ranked 14^th on the Top500. Clearly there have been significant improvements between the launches of these two machines.

In my last Green500 article, I included a chart like the one shown above. The point was to illustrate that it’s harder to be efficient as machines become larger (i.e., more compute capability). It is one thing to be efficient on a single chip where much of the communication is very local and it’s another to scale up to a machine that needs 10+MW and occupies its own room. The red line indicates how this efficiency drops off as the machines increase in size. This plot should also show what a leap forward the new Sunway TaihuLight machine is, not only in terms of performance but also energy efficiency. It is way off the line that the other supercomputing machines are currently on today and in a good way too. It is up and to the right in the direction of the Exascale goal posts marked as red and purple dots.

The PEZY Shoubu and Satsaki are the top two Green machines and are worthy in their own right of recognition. One thing they share in common with TaihuLight is that they also do not use a GPU Accelerator/Co-Processor. PEZY instead uses its own PEZY-SCnp. The next 20 machines on the Green list all use NVIDIA Tesla to help get their power and performance numbers.

In Olympic parlance, China takes home the gold and silver and the USA takes the bronze in the Top500 and on the Green list it’s Japan taking home the gold and silver and China the bronze.

It will be interesting to see if this is a new trend of custom designed chips upending the GPUs or if the newer GPUs will have an answer and be able to leapfrog back into the lead.