"Welcome to the Tokyo Institute of Technology. We are one of the national supercomputing centres in Japan. Currently we have the second fastest supercomputer in Japan. We currently are ranked 22 in the world. We have been as high as 4 back in 2010 when we had the first introduction of the Tsubame-2 machine on the list. Of course the machine is aged. As a research centre in supercomputing we have a number of groups, including mine, with tens of people that are working with our industrial partners and also with research labs such as those in DoE Livermore, Oak Ridge, and so forth.
In that we are trying to elevate our performance to the next generation. So we are planning for Tsubame-3. A renewing or update of our machine in 2016. The Tsubame-KFC machine that you see here is a prototype of the Tsubame-3. One of the most important properties is to basically work on power, because right now power is the limiting factor for how much capacity we can afford in a system. Especially in a university environment like ours we cannot afford at 20 MW data centre. At the normal operation it is about 1 MW and then at maximum we can afford only 2 MW.
So what do we do when the university doesn't give us more electricity? Basically we try to be at top-class despite the shortcoming in our facilities. So we really have to innovate in terms of our power performance. KFC is an attempt to try to minimise energy and the cooling power required. Be able to do smart energy management, be able to govern and manage the machine with job scheduling and so forth in order to obtain the maximum efficiency of the machine. And we have been very successful in this endeavour, such that we were ranked number one on the Green TOP500 back in 2013 and again in 2014.
On the TOP500 list that was announced just this morning, we were the most power efficient machine in the world, although the Green TOP500 list differs a little bit.
The technology we use here is in partnership with the green revolution calling. And as you can see, this duck is floating in oil, it is not being fried, which means that the whole temperature is not that high. It's not high enough to fry this duck, but still it is at about 35 to 45°C. So the server being immersed in the liquid itself will stay in that range.
Because the fluid, in this case, has so much thermal capacity, even with the immensely heat dissipation from the server's special chips, the temperature gradient remains fairly small.
It also allows the chips to be very cool. Each of these nodes have 4 GPUs, so it's a dense architecture. For each of the GPUs in a comparative air cooling configuration, the temperature will rise to something like 80°C, whereas immersed it is at 40 to 45°C, even at full load.
This allows for chips to operate at nominal temperature, and also be able to clock or voltage boost, when required, we even have higher performance, but we also keep the components' temperature low. This allows to have a much longer life expectancy, that is: we will keep the temperature operating at really low values.
The environment is under some shield. So there is no part touching oxygen. There is no corrosion going on. There are no moving parts.
We don't get any vibration. We don't know whether Tsubame-3 will actually use immersive cooling, but definitely hot water cooling. It will have other innovative features that will accommodate Cloud and Big Data features into the machine. Some of those features you may see in a Cloud have never been incorporated into a supercomputer. You will see for the first time in Tsubame-3 container technology, and various types of facilitation of massive use of machine learning algorithms.
So Tsubame-3 will again become a very innovative machine, not only high-performance, not only will it be a very high efficient machine, but it will also cover a broader range of applications that we have seen with Tsubame-2. And I hope that will set much higher standards for the HPC community, in order to strive for convergence with Big Data and Cloud infrastructures."