AMD details new power efficiency improvements, update on ’25×20′ project

[eXpLoReRs]

Energy efficiency is (if you’ll pardon the pun) a hot topic. Foundries and semiconductor manufacturers now trumpet their power saving initiatives with the same fervor they once reserved for clock speed improvements and performance improvements. AMD is no exception to this trend, and the company has just published a new white paper that details the work it’s doing as part of its ’25×20′ project, which intends to increase performance per watt by 25x within five years.

 

If you’ve followed our discussions on microprocessor trends and general power innovation, much of what the paper lays out will be familiar. The paper steps through hUMA (Heterogeneous Unified Memory Access) and the overall advantages of HSA, as well as the slowing rate of power improvements delivered strictly by foundry process shrinks. The most interesting area for our purposes is the additional information AMD is offering around Adaptive Voltage and Frequency Scaling, or AVFS. Most of these improvements are specific to Carrizo — the Carrizo-L platform doesn’t implement them.

 

AVFS vs DVFS

 

There are two primary methods of conserving power in a microprocessor — the aforementioned AVFS, and Dynamic Voltage and Frequency Scaling, or DVFS. Both AMD and Intel have made use of DVFS for over a decade. DVFS uses what’s called open-loop scaling. In this type of system, the CPU vendor determines the optimal voltage for the chip based on the target application and frequency. DVFS is not calibrated to any specific chips — instead, Intel, AMD, and other vendors create a statistical model that predicts what voltage level a chip that’s already verified as good will need to operate at a given frequency.

 

DVFS is always designed to incorporate a significant amount of overhead. A CPU’s operating temperature will affect its voltage requirements. And since AMD and Intel don’t know if any given SoC will be operating at 40C or 80C, they tweak the DVFS model to ensure a chip won’t destabilize. In practice, this means margins of 10-20% at any given point.