ARM Community: One Size Does Not Fit All for Mobile Systems - ARM Community

Last week I was in Korea and Taiwan as part of the ARM Technical Symposia series of events happening in November and December. With the momentum behind Mali GPUs I was asked to come and talk about the innovations that we have seen recently, the challenges that we see coming up and how we propose to help to address them. One of these key opportunities is round the potential being enabled by graphics processing. Mali GPUs can now be found in over 200 end devices and many of these are coming out from the vibrant Android smartphone and tablet markets.

Graphics are becoming more central to consumer devices and we have seen a step change in the graphics performance available in smartphones, tablets and DTVs in the last year. This increased appetite for high quality graphics creates challenges for the OEMs, silicon providers and IP providers, such as ARM. We see the pressures coming from increased resolutions, increased complexity and demand for faster more fluid interfaces. For mobile devices these challenges needed to be addressed within the tight limits of power consumption envelopes. ARM predicts that as the complexity grows, the pressure on the underlying system will require up to 35% energy efficiency savings to be made annually from the GPU and System IP on the same silicon process. This is extremely challenging and is a challenge for the industry as a whole and not just ARM, and is the main focus of the ARM team who are thinking and working differently to make this a reality. The main areas we are looking at are: micro-architecture changes in the GPU itself, running the right task on the right processor; manufacturing process optimization specific for graphics; as well as adding in new innovative features to increase the energy efficiency across the system.

My presentation not only looks at the GPU aspects, which are closest to me, but the importance of the system and how close linkage between the CPU and GPUs can bring new levels of efficiency that will be required to address the consumer needs, within the lower power budget required by mobile devices.

ARM is leading the market with innovative solutions to address the challenge of increasing performance whilst minimizing power consumption. Technologies such as big.LITTLE introduced the concept of pairing a Cortex-A15, high performance CPU designed for performance intensive tasks such as web browsing, with the Cortex-A7 designed for the always-on always-available type of tasks. This has recently been extended to cover the newly announced 64-bit ARM products, Cortex-A53 and Cortex-A57, which are focused on processing tasks in the most energy efficient way. ARM extends this task placement for efficiency further by being the first to have a Full Profile GPU Compute solution for mobile devices with the Mali-T604, now found in the Samsung Chromebook and Google Nexus 10. GPU Compute enables you to identify what tasks would be more efficiently done if handed by the GPU rather than the CPU.

Memory is another crucial point of efficiency optimization in a system – the less memory bandwidth you use the lower the power consumed. Memory bandwidth efficiencies have been a focus of the Mali architecture from the start and it continues to be fundamental to the latest Mali-450 MP and Mali-T600 series.

Doing things as they have always been done is not going to address these demands so ARM took an innovative approach to a potential large efficiency feature – compression. The solution is called Adaptive Scalable Texture Compression (ASTC) – which brings improved image quality whilst reducing the memory bandwidth and energy consumed. ARM is all about partnership so we worked with AMD to polish the ASTC concept and it has now been approved as an extension to Khronos OpenGL ES 3.0 API. Energy efficiency across the system is the answer.

ARM several years ago identified the need to optimize the physical IP to match the underlying process technology with the processor being manufactured. The result of this was the ARM POP IP products – these are core-hardening acceleration technology to enable the production of the best ARM processor implementations, in the shortest time to market. In October ARM announced a POP IP solution for Mali-T628 and Mali-T678 on TSMC 28nm HPM process technology – bringing the benefits that have been experienced in the Cortex CPU area to Mali GPUs. The specific benefits this POP IP can bring is up to 27% higher frequency, 24% lower area and 19% lower power compared to implementations not using the POP IP. This customized physical IP development seeks to enable the most efficient GPU implementations at process geometries of 28nm.

So we have come a long way with Mali GPUs over the last few years and we expect to see this pace to continue into and throughout 2013 but each stage brings its new challenges. I feel that ARMs combination of knowledge and experience from designing IP addressing markets, such as mobile, for over 20 years with the innovation to look at the system as a whole will enable ARM to address the upcoming challenges and beyond.

Related Blogs

• Mali-T600 Series Completing the ARM 64-bit System Story
• Mali-T600 series - The new Nexus between Graphics and Compute
• 1st Mali-T604 based device comes to market - Google Chromebook from Samsung
• How low can you go? Building low-power, low-bandwidth ARM Mali GPUs
• ARM Unveils Details of ASTC Texture Compression at HPG Conference - Part 1
• ARM Unveils Details of ASTC Texture Compression at HPG Conference - Part 2

Pete Hutton, General Manager of Media Processing Division, ARM.
Pete has almost 30 years experience in the industry with a background in processor design. As the GM of the Media Processing Division at ARM, he is responsible for the team covering Graphics and Video Technology. Prior to this, Pete was the VP of Technology & Systems at ARM’s Processor Division, where he was responsible for Advanced Product Development on new Processors (including big.LITTLE and v8), System Architecture Development, Open Source Tools and Software, Compute Subsystems and Physical Implementation.