The AppleTV Cluster

Energy efficient parallel computing on consumer electronic devices

Total power consumption and energy efficiency all components have become important considerations for the design of datacenters and high performance computing facilities. For the next generation of supercomputers (exascale systems), energy efficiency will even become the single most important constraining factor. On the other hand, consumer electronic and handheld devices have always been designed with energy efficiency in mind, with low-power ARM-based CPUs powering the majority of devices today.

Driven by uses cases such as HD streaming, real-time video encoding, virtual reality, and 3D gaming, there are strong market forces driving the future development of theses mobile devices towards more powerful compute capabilities. Dual and even quad core ARM designs with GPU integration have for example been announced for the next generation of tablet computers.

In the light of these developments, it is therefore an interesting question to ask if consumer electronic devices can be viable building blocks of HPC systems of the future. The supercomputing market is for the most part too small to sustain its own CPU ecosystem and in the past has leveraged workstation and server CPUs. In the future the underlying technology might very well come from consumer electronic devices.

With our AppleTV Cluster we try to provide a data point on the current state of things with regards to energy efficient parallel and distributed computing on ARM powered consumer electronic devices. The 2nd generation Apple TV (ATV2) shares its hardware internals with the iPad: Apple A4 processor (ARM Cortex-A8) running at 1Ghz, a PowerVR GPU, and 256MB RAM. The device is small (3.9 x 3.9 x 0.9 inches / 9.8 x 9.8 x 2.3 cm) inexpensive (100 USD / 120 EUR) and consumes only about 6 Watts. Our goal with this project is to build a small (currently 6 node) proof-of-concept cluster out of ATV2s.

We are currently evaluating the cluster with respect to its performance characteristics and limiting factors. We will also analyze the industry trajectory that we believe might make consumer electronic based technology viable building blocks for future HPC system designs.

Note: This project is work in progress. Please check out this whitepaper for some early results of our evaluation. For questions please send email to fuerling@nm.ifi.lmu.de

Acknowledgements: Thanks to Apple for producing some nifty hardware and to the jailbreak community for making it accessible.

AppleTV is a registered trademark of Apple, Inc.