ARM moves to combat Intel's Atom
ARM plans to go head-to-head with Intel with a revised version of its A9 processor core that the UK company claims is much faster or more power-efficient than the Atom that is used in many netbooks.
Instead of using low-power process technologies, which the company has targeted for its cellphone-oriented processors, ARM has chosen to aim for higher-power technologies such as TSMC’s 40nm general-purpose (40G) process with the Osprey, a dual-core version of the A9.
The normal problem with the general-purpose processes versus their low-power counterparts is that they leak far more current, although they can support more than double the clock rate. But, by designing circuits to power down when not in use and run operations faster, it is possible to counteract the effects of the higher leakage, which is what ARM has done with the Osprey.
“Using the Atom as a reference power profile, this offers five times the performance for the same power profile or five times less power for the same performance,” claimed John Goodacre, director of programme management at ARM. “It’s not the traditional place where ARM plays. This is a new space for us.”
Goodacre said the comparison was made using the CoreMark benchmark published by EEMBC, which allows the benchmark code to run on both cores of the 2GHz Osprey, against the single-core Intel Atom N270 running at 1.6GHz.
One of the big differences between the Atom and the Osprey in terms of power saving, said Goodacre, lies in the way that the on-chip level-two can be powered down more frequently on the ARM design. Because the level-one cache will be used 80 per cent of the time, accesses to the level-two store will typically be in bursts, making a shift into a low-leakage ‘retention’ mode viable. In contrast, the Atom’s cache will move into a power-saving mode when the processor is idle or sleeping.
Goodacre said chips based on the Osprey will probably come within a year to 18 months, following on from the first A9 processors based on process technologies such as TSMC’s 40nm low-power version. Although 40nm is the initial target, he said the new crop of high-k, metal-gate sub-40nm processes coming from the Common Platform partners as well as from TSMC will help drive performance up.
“High-k does wonders for the leakge. It allows you to push the design harder,” Goodacre claimed. “But 40G is only just coming to the mass market.”