Intel tera-fies hacks with vision of the future

Written by Wil Harris

September 25, 2006 // 10:43 p.m.

Tags: #cores #idf #intel #tera-scale

IDF Fall 06: Intel's Kevin Kahn updated assembled hacks on the state of the firm's leap to Tera-Scale computing.

No, that doesn't mean scary technology, or fit porn-star technology. It means tera as in giga as in mega - data an order of magnitude more than we operate with today.

Intel has already moved from single to dual core, and is about to make the leap to quad. It's also adamant that we will see 10s of cores on processors over the next couple of decades. Computing is going to be "Heavily parallel", according to Kahn, and this means lots of cores, lots of I/O capacity and lots of network capacity.

In terms of I/O, one of the major problems is getting all the cores to talk to each other effectively. On Kentsfield, Intel's quad-core part, cores talk to each other using the FSB. When you get up to masses of cores, that's not really effective, and Intel has been researching integrating communication into the processors, either using a mesh of communications channels that route all around the cores in a grid-like system, or a bi-directional ring bus that transports data around the cores linearly. Whilst mesh interconnects route traffic around bottlenecks and failures (and are faster), they are more complex and expensive.

Physically fitting that amount of cores onto processors gets difficult, even as process technologies scale down further than 65nm. Intel has been investigating core 'stacking', working out how to put them on top of each other.

One of the major problems to overcome is the interface between the processors and the peripherals - the I/O hub. This is especially important when you are talking about 100 gigabit LAN connections. Intel is proposing using core capacity on the CPU to integrate the network controller straight onto the chip, meaning lower latencies and faster communication.

Key to getting all this working is allowing the components to talk quickly enough. You can do a limited amount with copper interconnects, but Intel is working hard, Kev says, at getting optical communications working cheaply and quickly.

Let us know your thoughts over in the forums.


View this in the forums