ARM mulls 'big-little' scheme for graphics
Peter Clarke
10/20/2011 11:52 AM EDT
LONDON – As ARM Holdings plc rolled out its Cortex-A7 core and the "big-little" flexible power-performance scheme based upon it on Wednesday (Oct. 19), a company executive tipped that the company is also considering the application of a similar scheme to its graphics processor cores. Peter Hutton, general manager of the media processing division at ARM, said: "We are looking at a little-big approach for Mali."
At present the big-little scheme applies to the pairing of the A7-A15 cores and allows software to migrate between them based on the processing performance required. Meanwhile ARM's Mali T604 graphics processor supports the OpenCL parallel programming environment and the notion of applying the GPU to parallelizable general–purpose processing tasks. At present there is a deal of hand-coding and manual partitioning that has to be done to break out code that is suitable for running on a GPU.
Hutton was, until August involved in the design of the Cortex-A7 energy efficient processor and the big-little flexible processing technology.
The T604 includes four shader cores, each of which contains two arithmetic pipelines, one texturing pipeline, and one load/store unit. The four shaders share a coherent L2 cache, an MMU, a tiler, and a Job Manager. This latter block is a key component because the shaders are multithreaded. The Job Manager can dynamically move threads among the shaders. It can be seen that there is already power-performance scalability at the thread level inside the Mali T604.
In that regard one could consider the A7, A15 and Mali TXXX, which are likely to be implemented monolithically, as a set of resources that software should find a way to harness optimally based on a set of defined parameters, most notably minimum latency and minimum power consumption.