Folding at home (FAH) has taken the Guiness World Record for being the most powerful distributed computing network with a top speed of over 1 petaflop - (a thousand trillion calculations per second).
This is a remarkable achievement and shows the immense power that can be brought to bear by spare computing power used in a distributed network. The key here though is massive parallelism which means the various nodes in the network (your PC or PS3) are all doing different jobs at the same time and are at various points through these jobs. This is what made FAH and the old title holder Seti at Home (a search for extraterrestrial life) so scaleable.
Individual computers on the network download work units from the central repository, process them individually and then resubmit them back to the central core for post processing.
This is in contrast to say the Earth Simulator of Japan, a massive supercomputer capable of running huge simulations with ridiculous numbers of variables and calculations very quickly but where everything is interdependent. Likewise the ultimate aim of the BLUE project from IBM and the US Department of Energy is to be able to simulate all the forces and atoms of a nuclear explosion to simulate what's happening to USA's aging atomic weapons stockpile as they are no longer allowed to perform live tests.
This doesn't take anything away from their achievement, however it does go to show just how much wasted processing capacity there is lying around on the network.
The FAH project ramped up from 250 Teraflops (trillions of instructions per second) to just over a petaflop by the introduction of 670,000 PS3 owners supplying their hardware, up from the 200,000 PC users who got it to 250 Teraflops. Given that there are over 6 million PS3s in the wild this represents about 10% of the total Ps3 userbase - a quick calculation indicates that PS3 owners alone, should they all connect up to the internet, could provide about 7.5 Petaflops of processing power... this is beore we take into account PCs, XBoxes and Nintendo Wiis.
What this illustrates to me is that many of these projects are limited by their publicity and how "glamourous" they are. Taking nothing away from the geekiness of searching for ET or the importance of seeing how protein folding will affect drug development in the future, a more elegent solution would be an open framework that users subscribe to which is then used by anyone who wants to create a distributed processing application.
For the end user it is seamless and the for the multitude of public projects requiring raw processing cycles it gives them to opportunity to get access to larger numbers than their marketing budget would otherwise provide for. Even private companies could pay to rent processing time thus investing funds back into the project for ongoing development or optimisation.
Showing posts with label supercomputers. Show all posts
Showing posts with label supercomputers. Show all posts
Friday, 2 November 2007
Tuesday, 1 May 2007
super computer required to simulate half a mouse brain
Scientists have published that the've used the IBM Blue Gene L supercomputer to simulate half of a typical mouse's brain. More accurately they've simulated about half the neurons and just over half the number of synaptic connections for 10 seconds - which because the simulation was running at about a tenth of normal speed showed about 1 second's worth of realtime information.
You can read the whole story here
Don't get me wrong, the guys at Navada Uni have my utmost respect. I studied a lot about cognition and neural networks when I was at Uni, in fact I specialised in it with a degree in Computer Science and Psychology so I know how hard this is to do.
What gets me going though is the reasons behind doing it. As can be seen here, the top Supercomputer in the world can be brought to its knees by modelling half a mouse brain for a very limited period of time. The reason for this is the sheer number of connections [synapses] that occur between neurons - a single neuron in a mouse can influence the behaviour of about 8,000 other neurons. It doesn't take long for the cascade to build up and your computations to start slowing down.
What I find most interesting is that Blue Gene is designed to simulate molecular interactions particularly associated with the degradation of US Nuclear Weapons but it grinds to a halt with half a mouse brain.
I must say though that when I was playing with this over 10 years ago we were talking about ant or fruit fly brains which are merely hundreds of neurons in size and our computers were falling over. Given that baseline, the achievement these guys have made is incredible, although using the most powerful computer on the planet just shows you how far we are from modelling a human brain.
Human brains typically have about a 100 billion neurons with many thousands of synapses. Rough estimates put the number of connections at about a quadrillion synapses which for those of you that like zeros looks like this: 1,000,000,000,000,000
Also to note was that when this is done typically one uses random assignation for where the synapses end up, it isn't a true model of how a brain works as there would be too much information to configure and would have to be done by hand. In these models the neurons are loaded into the system then randomly assigned a number of dendrites which randomly point to other neurons. You don't get real behaviour as in hearing and vision and the like but you do get a sense of how the flow of stimulus and response works.
My own conclusion from my studies and keeping abreast of the topic since leaving formal education behind is that small neural networks specialised to a particular task are more likely to have results than large scale applications like this. Even mother nature adopted this process as you can see in evolutionary history that old structures are built upon by new, more specialised ones - you only need to look at a reptilian brain and compare it with our own, particularly the basal ganglia cluster to see the similarities in structure and function. In pulling these structures together you can then start achieving something that is greater than the sum of its parts.
You can read the whole story here
Don't get me wrong, the guys at Navada Uni have my utmost respect. I studied a lot about cognition and neural networks when I was at Uni, in fact I specialised in it with a degree in Computer Science and Psychology so I know how hard this is to do.
What gets me going though is the reasons behind doing it. As can be seen here, the top Supercomputer in the world can be brought to its knees by modelling half a mouse brain for a very limited period of time. The reason for this is the sheer number of connections [synapses] that occur between neurons - a single neuron in a mouse can influence the behaviour of about 8,000 other neurons. It doesn't take long for the cascade to build up and your computations to start slowing down.
What I find most interesting is that Blue Gene is designed to simulate molecular interactions particularly associated with the degradation of US Nuclear Weapons but it grinds to a halt with half a mouse brain.
I must say though that when I was playing with this over 10 years ago we were talking about ant or fruit fly brains which are merely hundreds of neurons in size and our computers were falling over. Given that baseline, the achievement these guys have made is incredible, although using the most powerful computer on the planet just shows you how far we are from modelling a human brain.
Human brains typically have about a 100 billion neurons with many thousands of synapses. Rough estimates put the number of connections at about a quadrillion synapses which for those of you that like zeros looks like this: 1,000,000,000,000,000
Also to note was that when this is done typically one uses random assignation for where the synapses end up, it isn't a true model of how a brain works as there would be too much information to configure and would have to be done by hand. In these models the neurons are loaded into the system then randomly assigned a number of dendrites which randomly point to other neurons. You don't get real behaviour as in hearing and vision and the like but you do get a sense of how the flow of stimulus and response works.
My own conclusion from my studies and keeping abreast of the topic since leaving formal education behind is that small neural networks specialised to a particular task are more likely to have results than large scale applications like this. Even mother nature adopted this process as you can see in evolutionary history that old structures are built upon by new, more specialised ones - you only need to look at a reptilian brain and compare it with our own, particularly the basal ganglia cluster to see the similarities in structure and function. In pulling these structures together you can then start achieving something that is greater than the sum of its parts.
Subscribe to:
Posts (Atom)
