Computational sprinting is a method that is used to increase the performance of a device by consuming more power. This method is particularly useful for applications like data analytics that require short bursts of computation.
How does Computational Sprinting Work?
In computational sprinting, many independent chip multiprocessors share a power supply. However, sprints are limited by the thermal limits of the chips and the power limits of the rack. Thus, the approach uses powered-down cores that are activated for split-second, intensive parallel computations. In other words, a chip temporarily exceeds its sustainable thermal power budget to achieve instantaneous throughput, after which the chip must return to nominal operation to cool down.
The Computational Sprinting Game
The computational sprinting game is a multi-agent perspective for managing sprints. In this game, strategic agents decide whether to run a sprint based on application phases and system constraints. The game produces an equilibrium that improves task throughput for data analytics workloads by 4 to 6 times compared to previous greedy heuristics, and is within 90% of an upper bound for the throughput of a globally optimized strategy.
The Benefits of Computational Sprinting
While transistor density continues to increase, voltage scaling has stalled, so power density increases with each generation of technology. Particularly in mobile devices with limited cooling capabilities, these trends are leading to a utilization limit where sustained chip performance is primarily limited by power rather than area. However, computational sprinting allows devices to achieve higher instantaneous throughput and improve responsiveness for applications that involve short bursts of computation in response to user activity.
Frequently Asked Questions (FAQ)
- What is computational sprinting?
- How does computational sprinting work?
- What is the computational sprinting game?
- What are the benefits of computational sprinting?
Computational sprinting is a method that is used to increase the performance of a device by consuming more power.
In computational sprinting, powered-down cores are activated for split-second, intensive parallel computations. A chip temporarily exceeds its sustainable thermal power budget to achieve instantaneous throughput, after which the chip must return to nominal operation to cool down.
The computational sprinting game is a multi-agent perspective for managing sprints. It produces an equilibrium that improves task throughput for data analytics workloads by 4 to 6 times compared to previous greedy heuristics.
Computational sprinting allows devices to achieve higher instantaneous throughput and improve responsiveness for applications that involve short bursts of computation in response to user activity.
The Bottom Line
Computational sprinting is a useful method that can significantly improve the performance of devices for data analytics and other applications that require short bursts of computation. By activating powered-down cores for split-second, intensive parallel computations, devices can achieve higher instantaneous throughput and improve responsiveness. The computational sprinting game is a useful tool that helps in managing sprints for optimal performance.
