Latest Top500 list, looking beyond the number 1

Author: Adrian Jackson
Posted: 21 Jun 2016 | 17:13

There's been a lot of discussion about the latest Top500 list, released this week at ISC16.  Most of the interest has been in the whopping new Chinese system, Sunway TaihuLight, which has come in at number 1 on the list with a massive 93 PFlop/s rpeak Linpack performance, and 125 PFlop/s rmax theoretical peak performance (3 times bigger than the previous number 1 system).Top500

Whilst this is a very interesting system, and much bigger than is currently planned elsewhere, it's not unknown for very large systems to come in and dominate the list like this.  Back in 2002, the Japanese Earth Simulator system became the number 1 machine with an rpeak of ~5x that of the previous number 1 system, and it stayed as the top machine for a number of years.

There are plenty of good blog posts/discussion about TaihuLight out there (for example; Glen Lockwood, Jack Dongarra, Karl Rupp) so I'm not going to focus on that system. However, there are some other interesting trends in the latest list that are worth a brief look.

Vendors

Vendor system share, Top 500 list, November 2015Vendor breakdown of Top500 list (November 2015)

The first set of data that we can look at is the breakdown of vendors in the list (all the graphs in this post are generated from the Top500 site). We can see that going from the November 2015 to June 2016 lists, Lenovo has taken a much bigger share of the total number of systems. It's always surprising to see HP, Lenovo, etc as large players in the Top500, but it's worth bearing in mind that whilst a lot of focus is on the large systems, the majority of the Top500 list is much smaller scale systems. Lenovo seems to have picked up a big share of these recently, to the detriment of HP/SGI/IBM (the IBM part is to be expected since they exited the HPC market and aren't yet in a position to re-enter).

Vendor share of the Top500 list, June 2016Vendor breakdown of Top500 list (June 2016)

Cray's performance share has suffered a bit, but this is likely to be because of the very large number 1 system skewing the results, rather than a large drop off in Cray performance.

Countries

Looking now at country data, we can see a very significant shift between the two lists, with China and the USA reversing positions in both number of systems and performance. China has gone from having ~20% the performance share to ~40% (not surprising since the number 1 system is so large), but more importantly has gone from having half as many systems as the US in the list to having the same percentage.

Counties system share, Top 500 list, November 2015Counties breakdown of Top500 list (November 2015)

Now this could be an actual change, or just a difference in behaviour (people chosing to enter their systems for the Top500 list or not), but either way it's a big change around in 6 months. It's important to remember that the Top500 list isn't a factual list of HPC machines in countries, but a self-selecting/self-publicising list of people who want to enter their system. There may be many reasons to not enter a system: because it requires effort and expense (runtime on the system), because the system is secret, or because you don't agree with ranking systems (ie Blue waters).

Countries system share, Top 500 list, June 2016Counties breakdown of Top500 list (June 2016)

It's also worth noting that the UK is steadily dropping down the list, both in number of systems and performance share. It's not so long ago that the UK was in second or third spot in terms of numbers of systems, and high up in the performance list, but we're no longer a big player in large-scale HPC systems. This could simply be a timing issue, for instance the Met Office will shortly have a large HPC system (likely top 10 on the current list) but it's not quite on the ground yet.

Processors

Finally, it's interesting to look at breakdown by processor.  Here there is much less change evident. The percentage share of systems with Intel processors hasn't really changed at all, although Haswell systems feature much more, with Ivybridge becoming less important and Broadwell not yet established. 

Processor share of the Top 500 list, November 2015Processor breakdown of Top500 list (November 2015)

In terms of performance, the Intel share seems to have gone from ~75% to ~65%, so again, given the size of the top 2 machines in the list, this isn't a big blow to Intel. They are still massively dominant in HPC systems. Indeed, the AMD share has decreased further between the two lists, so hopefully we'll see some new products from AMD shortly to challenge this Intel domination and provide some competition in the market.

Processor share of the Top 500 list, June 2016Processor breakdown of Top500 list (June 2016)

The new processor(s) developed by the Chinese may one day become important for global HPC systems, but currently we're not seeing any impact from them.  Chinese vendors, on the other hand, are starting to make in-roads into the market, so there is certainly competition for packaging and selling systems. It's likely that the November 2016 list will see significant change again, with large-scale KNL systems making an impact, and maybe even some Pascal-based machines.

It's certainly interesting times in the HPC market, but maybe not as interesting, yet, as the news would suggest.

This post also appears on our Medium.com channel.

Author

Adrian Jackson, EPCC
You can often find Adrian on Twitter.