Page 9

EETE JULAUG 2013

Multicore challenge conference Multicore systems face the tools challenge By Nick Flaherty The move to more heterogeneous multicore systems is going to fundamentally change the way code is developed in mobile, says a leading IP supplier. “This is the most exciting decade I have seen for computing,” said Tony King-Smith, executive vice president of marketing at Imagination Technologies, talking at the Multicore Challenge Conference run by verification expert TVS. “There are opportunities to lead this global change here in the UK. He points to the increasing split in the mobile market between ARM and other processor architectures all running different versions of Google’s open source Android operating system. Intel has started to have some success with x86-based Atom processors in mobile handsets and tablets, while Imagination has acquired the MIPS processor line and is combining this with its PowerVR graphics and video technology and its programmable radio front end. Coupled with the move to technologies such as ray tracing, this creates a widely varying set of requirements, he says. “We have three architectures in the Android world but the apps have to be able to traverse CPUs in the same way they do across GPUs and radios. We have to break this dependency on the CPU instruction set architecture and this is part of the future for heterogeneous processing,” said King-Smith. “Google is well aware of this problem,” he said. There are different ways to tackle this issue, he says, including new capabilities in the LLVM tool chain (see box). “We are quite excited about LLVM and the portable binary format with binary translation,” he said. Another exciting approach actually changes the way systems are developed. Instead of downloading an app for a particular ISA, a generic app is downloaded that investigates what hardware resources are available. This ‘discovery’ app could be written with LLVM’s binary translation capability or a higher level language such as Java. Once the discovery app determines the hardware available, it downloads different optimized blocks for the different hardware elements, creating the optimal software. This is not simple to do. “Discovery will be a fertile area for research and innovation,” said King-Smith. More importantly it also changes the way the software is developed, he says. Instead of starting off with the data structures, you start with the discovery app. That provides the base software that is already available, and the developer then fills in the gaps in the software ecosystem, concentrating on the added value rather than re-inventing software for multiple different platforms. Putting all this together with the right balance of memory and performance determines the power consumption and performance of the system. “This is a very key area,” he said. “Attention to detail makes a huge difference. I’ve seen four to five times the performance difference using exactly the same set of IP, so getting the balance is utterly key in making these systems work.” The tool issue is the main reason why multicore computing has been slow to take off, says Prof David May at the University of Bristol. He points to homogeneous architectures where the programming is significantly easier. This was the approach he took with the multicore system used by XMOS Semiconductor where he is a founder. Fig. 1: Imagination Technologies sees a discovery tool that looks at the hardware and downloads the relevant optimized code to build the code base for heterogeneous multicore devices. www.electronics-eetimes.com Electronic Engineering Times Europe July/August 2013 9


EETE JULAUG 2013
To see the actual publication please follow the link above