Andrew Dean is HPC Business Development Manager at OCF in the UK.
In this special guest feature, Andrew Dean from OCF writes that China has plans for building an Exascale prototype by the end of this year.
Back in 2013 I wrote the following blog expressing my opinion that I doubted we would reach Exascale before 2020. However, recently it was announced that the world’s first Exascale supercomputer prototype will be ready by the end of 2017 (recently pushed back to early 2018), created by the Chinese. I did some digging and wanted to share my thoughts on the news.
Looking at the most recent TOP500 list, interestingly, Exascale is still quite a big step away from the current largest system on the list, the Chinese-made Sunway TaihuLight. An Exascale system would have to be ten times the size of that current system, which is a massive jump technology-wise.
If you take a look at the graph, below, which tracks the performance development over the last 20 years of the TOP500, it is pretty much a straight line against it’s logarithmic scale – and has been this way since the 90s with a projected number 1 system reaching exascale around 2019. So, it would go against the trend of the last 20 years for an Exascale system to exist at this point in time as there are very few sharp peaks so far.
The largest peak/anomaly so far being in 2002 when the Japan Agency for Marine and Technology -Earth Science Earth Simulator system overtook the Lawrence Livermore National Laboratory – ASCI White, SP Power3 375 MHz system by approx. 5x Rmax performance. If a similar 5x ‘jump’ were to happen today this would still only get us (almost) half way to Exascale showing what step change this would represent. It’s also interesting to note this Earth Science Earth Simulator system held the number 1 crown until Nov 2004, again reinforcing how big a leap this 5x increase was with it taking so long for others to catch up.
I would be very interested to know what this prototype will look like: are we talking hardware, or simply part of a software stack? Due to the secrecy on the Chinese side, it is difficult to know what they are physically building, for example, the prototype may be significantly smaller than a normal cluster.
It would be great have more information to see what this type of system would be used for – and what kind of applications can run on it. That’s part of the problem with Exascale – we know it is possible to have that much processing power, but how you actually use it is a whole other story.
The vast majority of our customers have very varied workloads and are using tens, if not hundreds of applications on our clusters. For us, it’s not about getting one application to work well on one cluster – it’s getting them all to work. That’s why we’re seeing the continued dominance of traditional x86 two-socket socket systems, despite other interesting technologies being available such as IBM Power, ARM, FPGA’s, GPU’s, Co-processors – customers can run lots of things very well instead of just some things, and that is our focus.
Of course, it could be that the Chinese are building the Exascale model for the sake of it – so that they are the first to reach the target, like being the first man on the moon. The entire project will take a huge amount of money and resources, and in some cases, it would make more sense to have ten systems a tenth of the size that are more traditional architecture rather than one Exascale supercomputer.
I mentioned in my previous blog that costs, power and cooling would be the big challenges in terms of the build up to Exascale. I still think this is true, but should also add networking the system together, the general management of the supercomputer, and writing the applications to run on it to that list.
In summary, I’ll be really interested to see how the prototype Exascale system works when – or if – it is completed in late 2017. I’d also like to find out if the Chinese will keep the technology and specs a secret or if they take this opportunity to make the leap into the market with predominantly US players, like Intel and AMD. It could go one way or the other – as even though the Chinese sell plenty of servers, many of them are sold with non-Chinese chip technology. It will be really interesting to see a new player potentially added to the mix.
If this is a man-on-the-moon kind of project for the Chinese, I would expect they’ll use a CPU of Chinese manufacture (and maybe design, although if history holds true, they’re more likely to “resuse” an Intel design). The machine doesn’t have to be useful. It’s not clear how much use the Tianhe’s get.