Technology Review - Published By MIT
Log in to My.TechnologyReview.com | Register
Advertisement
[1] 2 Next »

Monday, August 20, 2007

A New Design for Computer Chips

An MIT spinoff introduces the first commercial chip with a mesh architecture.

By Kate Greene

smaller text tool iconmedium text tool iconlarger text tool icon
Multicore mesh: This chip, which measures about 40 millimeters square, contains 64 processors--or “cores”--connected to each other in a mesh network, a new chip architecture. The processor is currently shipping to companies that make videoconferencing technology and network routers.
Credit: Tilera

Today, MIT spinoff Tilera announced that it's shipping a computer chip with 64 separate processors whose design differs drastically from that of the chips found in today's computers. The new chip, called Tile64, avoids some of the speed bottlenecks inherent in today's chip architecture, and it can operate at much lower power, says Anant Agarwal, founder and chief technology officer of Tilera, based in Santa Clara, CA. Initially, Tile64 will be used in video applications such as videoconferencing systems, and in network hardware that monitors traffic to reduce e-mail spam and viruses.

Chips with multiple processing units, or "cores," are nothing new. But by allowing the cores to communicate directly with each other, Tilera has addressed a widespread concern about the viability of adding more cores to microprocessors. "Every processor in the market today is a multicore," says Agarwal. "The hope of the industry is to double the number of cores every 18 months. My prediction is, by 2014, we will have 1,000-core architectures. But the problem is, [current] architectures don't scale."

In existing multicore chips, each core communicates with the others via a set of wires called a bus. Performance doesn't necessarily suffer when two or four cores share a bus, but when 16 or more cores try to use it simultaneously, data can get backed up. Agarwal explains that Tilera's chip has no central bus. Instead, each core is connected to all the others. Also on each core is a full-featured processor, which can run an operating system, and memory caches, which hold data that needs to be quickly accessed.

In effect, the Tile64 has a mesh structure that's similar to that of the Internet, a network in which there are many decentralized nodes. One reason the Internet is able to pass around data so quickly is that packets of information are sent through a vast network and can avoid traffic jams. If everyone's e-mail had to go through a central server, there would undoubtedly be delays. Tilera's microprocessor, says Agarwal, "is very much like the Internet on a chip." And like the Internet, Tilera's chip can be scaled up gracefully; it doesn't need to be redesigned each time new cores are added.

The idea of using mesh architecture for multicore chips has been explored for at least a decade, in research labs at MIT, Stanford, and the University of Texas, Austin. And recently, Intel announced a prototype 80-core chip based on a mesh. But Tilera is the first company to offer a product that uses the new architecture.

"Having a lot of cores is good, but they must be able to communicate with each other at high data rates," says Jerry Bautista, codirector of Intel's terascale-computing research program. "There are advantages to using a mesh ... You can deal with traffic jams pretty easily." Bautista says that Intel researchers are trying to find the best way to implement mesh architectures--among other experimental designs--in future chips. But he also cautions that making massively multicore systems work efficiently isn't as simple as redesigning the hardware.

[1] 2 Next »

Comments

  • Software Model
    eightwings on 08/27/2007 at 10:43 PM
    Posts:
    2
    Avg Rating:
    5/5
    "Intel's Bautista says the marketplace may be ready for a chip with more computing power, but it would need to be low power and easily programmed."

    Multicore CPUs are not easily programmed and that's the problem that neither Intel nor Tilera has adequately addressed. And it's a problem that must be solved before the industry can see a massive switch to chips like Tilera's.

    In my opinion, the reason that current multicore processors are hard to program is that they were created for the wrong software model. The algorithmic model has been with us since the days of Charles Babbage and Lady Ada. It is not well suited for parallel processing because it is inherently sequential. It is time that we switch to a non-algorithmic model, one which is implicitly parallel. Sequential ordering should be explicit, not implicit as it is now. Project COSA is a step in the right direction.

    http://www.rebelscience.org/Cosas/COSA.htm

    The multicore CPU manufacturers have only themselves to blame. You don't design a new CPU architecture and expect the software model to change to accomodate it. It should be the other way around. The software model must come first.
    Rate this comment: 12345
Advertisement

Current Issue

Technology Review July/August 2008
The Business of Social Networks
The future of the Web is social. But can social-networking sites ever make money?
•  Subscribe
Save 41%
•  Table of Contents
•  MIT News

Magazine Services

Career Resources

MIT Technology Insider

Stories and breaking news from inside MIT about the latest research, innovations, and startups--in a convenient monthly e-newsletter. Subscribe today
Advertisement

More Technology News from Forbes

Advertisement
Advertisement
Advertisement
TECHNOLOGY RESOURCES
Advertisement
MIT Massachusetts Institute of Technology