Researchers Explore Alternatives To Dark Silicon

One of the features that AMD, Intel, Nvidia, and ARM have all pushed in the last six years is the evolution of their various power saving technologies. This allows chips to draw much less power, but it leaves increasingly large amounts of the processor unpowered the majority of the time. As process technologies shrink, this so-called 'dark silicon' proportionally increases; it's estimated that mobile / ultra-mobile processors manufactured at 11nm might only be able to power nine percent of their total transistors at any given time.

The question researchers have been asking is whether or not there's anything else the die might do that could increase application performance without creating additional thermal issues. Researchers at the University of California, San Diego, have an idea for how to simultaneously improve power efficiency and performance, dubbed GreenDroid. According to Michael Taylor, co-leader of the GreenDroid project, mobile chips need a fresh approach.

GreenDroid chips surround a traditional core with up to 120 ultra-specialized micro-cores (our term). Each of these cores is hard-wired to execute specific bits of code. Checking email, opening one's MP3 player, launching a browser, or running Java / Flash are all examples of functions that could be hard-coded into the processor (at least to some degree). Taylor states: "If you fill the chip with highly specialized cores, then the fraction of the chip that is lit up at one time can be the most energy efficient for that particular task."

An example of a 'conservation core.'

The San Diego team has already extensively modeled Android's performance and the needs of various applications and designed a microprocessor that should be capable of hardware-accelerating ~ 70 percent of Android code; it's sending its design out for fabrication in June. The ability to model such changes before silicon is built lowers the chance that the hardware coming off the fab lines could be subtly incompatible with the software written to use it.

It's unlikely that we'll see anything similar in the desktop or server market. It's relatively easy to anticipate the desires of a mobile user base, but PCs and servers run too broad a variety of programs to be easily bracketed and have much larger power budgets to play with. It's also true that manufacturers might be leery of rapid update cycles that leave phones with hardware accelerated features updated with software that can no longer take full advantage of their integrated micro-cores.

Even with these caveats, the mobile industry will likely be curious to see what sort of results the GreenDroid researchers measure on actual silicon. The mobile market is churning with multiple vendors, each eager for a leg up and a way to claim that their own devices offer the best battery life, fastest gaming/web browsing, and highest quality video. As die sizes shrink, handheld/ultra-mobile developers may well turn to such methods to continue pushing features while driving down prices.