by Paul Rudo on 31/05/12 at 5:03 pm
It’s no secret that power consumption has become a major sticking point for datacenter operations within the past few years. As processors become smaller and cheaper, they also require significantly more power and generate much more heat.
Today’s tightly-packed blade servers run significantly more heat into a much smaller area than previous generation of boxes. Today’s datacenters consume significantly more electricity per square foot, and the heat they produced must also be controlled by power-consuming cooling systems.
Of course, a smaller footprint hasn’t reduced our need for hardware. Cheaper hardware has only increased our appetite and presented opportunities for IT to support strategy in ways that were previously economically impractical.
As a result of this increasing usage and lower hardware costs, power consumption has suddenly emerged as a significant part of overall TCO for datacenters. This has spawned the growth of the “green computing” trend, where a new industry has emerged in order to help companies cut power-related costs and make IT services more sustainable.
One recent proposed solution proposed has been the introduction of “inexact” chips.
Since 2003, a number of universities – including Rice University, Switzerland’s Center For Electronics and Microtechnology, and Nyang Technological University – have been working jointly to develop new approximate computation technologies. This work was recently revealed at the ACM International Conference on Computing Frontiers, and met with very positive reactions.
The team used a video demonstration to show how chips could become 15 times more efficient in terms of space, energy and speed by allowing only a small tolerance for error. This efficiency was achieved using a few interesting mathematical methods.
- By carefully analyzing the probability of error for certain types of calculations, such as addition and multiplication, you can allow for errors in a way that minimizes impact on processes.
- Another approach to improving efficiency would be to prune portions of a digital circuit which are rarely used.
- Another more complex method called “confined voltage scaling” makes a few performance compromises in order to improve overall processing speed.
Inexact processing seems to show some very promising results when it comes to reducing expensive power consumption for servers. By simply pruning low-value parts of traditional processor, it was shown that size and power consumption could be cut down by half, while also doubling overall performance of the chip. And this benefit would only cause a very tiny drop in accuracy.
Some are saying that future computers would have separate processors which divide work into “exact” and “inexact” tasks, where low-precision processes could be routed through more efficient hardware to free up resources. This would be similar to how current computers have a separate video card to handle graphics.
Of course, the idea of reducing accuracy in order to gain a performance advantage is nothing new within software engineering.
Without the introduction of inexact “lossy” video and image formats such as JPEG, GIF, PNG and MPEG, the Internet would’ve taken at least 10 extra years to become widely commercially viable. And online video would still be a novelty, since bandwidth is just now starting to approach a point where it could potentially handle large lossless video files.
Organizations also use slightly inaccurate algorithms to allow rapid analysis of very large pools of real-time data through the use of computing clusters. Although inexact algorithms would be poorly suited for financial applications such as bank transactions, a slight drop in accuracy for Doppler weather analysis or genetic simulations might be perfectly acceptable if the trade-off were significantly reduced cost and a large increase in power or speed.
The inexact chip technology proposed could also potentially have a major worldwide social impact by making low-cost computers accessible to poorer nations. It’s no surprise that many developing countries – India in particular – plan to provide poor children with low-cost computers that could provide better access to education. In order for this to be practical, these machines would need to provide high-quality computing in the $50-$100 range. And inexact chip technology could potentially offer a way to reach this objective.