Tuesday, June 29, 2010

The future of computing? The SpiNNaker million processor computer

The future of computing is a very big claim. However if anything the future of computing does not lie in a common time, unified shared memory system - this pretty well describes every multicore computer made at the moment. The trouble is once you go to more than a few thousand cores the shared memory - shared time concept falls to bits. It just doesn't scale. Where then should we look for a new model other than biology? The brains of living creatures are capable of phenomenal processing power and yet have almost none of the features of the computers we build today. Don't you find this odd?

Lets look at how brains do it:

  • Does each neuron have a sense of the passage of time? Not in the traditional sense and maybe not at all. If it does it will probably only be only vague ordering.

  • Does a neuron share a memory space with every other neuron? No. The closest concept is an area effect of diffusion of chemicals which affect neighbours (which may not be synaptically connected) but that is very far from a shared memory concept.

  • Is each neuron connected to every other neuron? They have many connections to close neurons and some connections to neurons further away best described by a statistical distribution but otherwise, again, no.

So clearly modern computing bears as relation to a brain as sudoku does to quantum mechanics. The closest we get are the datacenters of a company such as Google but they enjoy far too much connectivity to be a good model.

Enter the world of academia because no commercial company would be stupid enough to commit the resources we can into a project like this (and all the ones before it which make it possible) - it is unlike anything that has come before and it is risky (i.e. it may not work!).

Project SpiNNaker

This revolutionary idea is nothing less than a plan to put together a computer of approximately 1 million cores with no common clock or shared memory and which can route messages with a model that approximates a neurological system.

Born in the University of Manchester's Advanced Processor Technologies Group the SpiNNaker project has taken shape in collaboration with the University of Southampton's School of Electronics and Computer Science (where I work), the Engineering and Physical Science Research Council and two enlightened commercial companies: ARM (for the processor IP) and SilĂ­stix (for their Network on Chip expertise).

SpiNNaker is a child of many parents, each a vital step on the path to it's genesis:
  1. AsipIDE GALS Design and Co-Simulation Framework - A hardware/software co-design and debugging framwork
  2. Transactional Memory - A new locking strategy which massively simplifies the ability to lock data to pass information between processes
  3. TERAFLUX: Exploiting Dataflow Parallelism in Teradevice Computing - A pan-european project looking at computers with massive numbers of cores
  4. The Balsa Asynchronous Synthesis System - A language and compiler targetting asynchronous, handshake driven logic design
The main centrepiece of the SpiNNaker project is a special System on Chip, the SpiNNaker SoC. Inside are 18 processors asynchronously connected via a blindingly fast network-on-chip and communicating with the other SoCs via hundreds of megabit links to the other SpiNNaker SoCs. With these SoCs it will take only 56,000 chips to reach the target of 1 million. The cores are ARM968 series processors capable of significant independent computation while communicating with their neighbours or the 1GBbit DDR SDRAM available to each SoC. Even within a SoC the cores do not share a common clock and communicate by passing messages which are routed to each other or outside to another SpiNNaker SoC.

A seriously ambitious project you say? Absolutely right. Vapourware or a dream? Hell no! Feast your eyes on this:




This is the first generation of SpiNNaker SoCs on a test board - there are four dual processor SoCs on this board each next to its accompanying 1Gbit ram chip (click on the picture for a very high resolution version). This board already holds 8 processors asynchronously interconnected with all the necessary debug hardware to perfect the design of the next generation.

Who ever said British Science was dull?

BIG DISCLAIMER: I am not one of the great minds trying to change the world with this project. At best I have helped a couple of people around the edges. I am a very big fan, however!

Laird Tpcm 7250 is as good as Honeywell PTM7950 as thermal paste / interface for PC

[This is not very scientific, however it is notable. At 7.5W/m-K vs the installed SYY-157 at 15.7 W/m-K it performed better in real world lo...