Computational sprinting

Abstract

Although transistor density continues to increase, voltage scaling has stalled and thus power density is increasing each technology generation. Particularly in mobile devices, which have limited cooling options, these trends lead to a utilization wall in which sustained chip performance is limited primarily by power rather than area. However, many mobile applications do not demand sustained performance; rather they comprise short bursts of computation in response to sporadic user activity.To improve responsiveness for such applications, this paper explores activating otherwise powered-down cores for sub-second bursts of intense parallel computation. The approach exploits the concept of computational sprinting, in which a chip temporarily exceeds its sustainable thermal power budget to provide instantaneous throughput, after which the chip must return to nominal operation to cool down. To demonstrate the feasibility of this approach, we analyze the thermal and electrical characteristics of a smart-phone-like system that nominally operates a single core (∼1W peak), but can sprint with up to 16 cores for hundreds of milliseconds. We describe a thermal design that incorporates phase-change materials to provide thermal capacitance to enable such sprints. We analyze image recognition kernels to show that parallel sprinting has the potential to achieve the task response time of a 16W chip within the thermal constraints of a 1W mobile platform.

Computational sprinting on a hardware/software testbed

“…Perhaps counter-intuitively, however, we find sprinting can actually result in net gains in energy efficiency by (i) amortizing the fixed uncore power consumption over a larger number of active cores and (ii) capturing race-to-idle effects. Our conclusions stand in contrast to our previously published predictions from simulation-based analysis of computational sprinting [45], which neglected uncore power and hence suggested that sprinting might at best be energy neutral.…”

“…A complementary approach to increase sprinting effectiveness is to engineer a system that supports longer sprints by including more thermal capacitance. Instead of leveraging only the specific heat of conventional materials, prior work proposed using the latent heat of a phase change material (PCM) to add thermal capacitance to a sprinting system [45]. Latent heat has the advantage that it can absorb substantial heat without a change in temperature.…”

“…Previously, we explored the feasibility of computational sprints that exceed sustainable thermal limits by an order of magnitude through a combination of modeling and simulation [45]. In this paper, we describe and measure a concrete implementation of computational sprinting in a hardware/software testbed that supports both parallel and frequency sprinting, offering a maximum sprint intensity that exceeds sustainable power levels by a factor of five.…”

Making JavaScript Better by Making It Even Slower

“…This analysis makes sense for some workloads and tasks, for example, the image recognition kernel in [13], and perhaps the UI controls on the platform (Section VI-E), but it may not capture the event-driven nature of JavaScript execution for a web page or application. Here, the workload is likely to be continuous, and in many cases, users may not be concerned with the speed at which individual tasks complete, or even if they complete.…”

“…The implication is that given fixed tasks, the best strategy is to execute them with as high a rate as possible. Computational sprinting [13], where execution rate is raised even beyond sustainable thermal limits for brief periods of time, is the extreme example of this "race to finish" approach.…”