While at the IEEE High Performance Extreme Computing (HPEC) conference in Massachusetts this week, Piotr Luszczek who serves as the Research Director for the University of Tennessee was talking about the ARM landscape and embedded LINPACK benchmarking. The BoF presentation was entitled "Anatomy of a Globally Recursive Embedded LINPACK Benchmark."
The presentation's abstract was:
We present a complete bottom-up implementation of an embedded LINPACK benchmark on iPad 2. We use a novel formulation of a recursive LU factorization that is recursive and parallel at the global scope. We be believe our new algorithm presents an alternative to existing linear algebra parallelization techniques such as master-worker and DAG-based approaches. We show a assembly API that allows us a much higher level of abstraction and provides rapid code development within the confines of mobile device SDK. We use performance modeling to help with the limitation of the device and the limited access to device from the development environment not geared for HPC application tuningLuszczek has uploaded his slides to the UTK web-site, but what's interesting for the common Phoronix reader are just his Apple iPad 2 test results.
The researcher mentioned that Apple introduced the Accelerate Framework to iOS4 and that they were to include an optimized ATLAS (Automatically Tuned Linear Algebra Software) library inside this iOS framework, just like with it shipping inside the Mac OS X framework. However, this framework didn't really work and there is no ATLAS for the iPhone/iPod/iPad so he took things a lot further on his own.
Through a long process, Piotr Luszczek was able to learn more about the iPad hardware itself through nano and micro benchmarking. After this, he was able to create an optimized algorithm by writing a Python script to generate various Assembly routines to test each one for the most efficient performance. A few ARM tweaks also got tossed into Atlas BLAS.
When benchmarking the Apple iPad 2, the University of Tennessee employee achieved 4 GFLOPS per Watt on the ARM SoC (measured at the chip level). As the below chart shows, he found the iPad 2 to be as fast as the Cray 2 super-computer from the 1980s. The Cray-2 was originally the fastest super-computer in the world when it originally premiered. Piotr Luszczek also found that the original iPad was about the speed of the original Cray-1 super-computer. The latest iPad meanwhile is just a small performance bump over the iPad 2.
It will still be a long, long time though before any tablet can be as fast as ASCI Red/White, Sequoia, or the modern Cray Jaguar that has held a LINPACK performance record.
The other interesting slide were some numbers on various devices for their power and performance.
If you are interested in maxing out ARM performance, see the Phoronix 12-core ARM cluster and then the 96-core ARMv7 solar-powered super-computer that was assembled out at MIT earlier this summer. Coming up in a matter of weeks will be the long-awaited many-core Calxeda ARM benchmarks premiering on Phoronix.