ARM has upped its game against Imagination Technologies, its rival in the field, unveiling a 16-core design in its Mali family of GPU (graphics processing unit) cores.
As mobile applications become more graphics-intensive and ARM also sets its sights on high end infrastructure, it needs to keep up with the GPU demands of its customers. A future Mali-T760 design will boast up to 16 cores, with higher performance and a 400% reduction in power consumption. Although the IP may not appear in commercial products for about a year, it already has several lead licensees, including LG, MediaTek and Rockchip, and production quality software will be available in three months’ time.
The T760 delivers up to 326GFlops and is designed to accelerate ARM’s recent gains in the mobile graphics market, where it has been narrowing the gap with Imagination and Qualcomm. The company says that Mali is now present in 50% of Android tablets and 20% of Android smartphones, but it is excluded from the iDevices as Apple is a customer of, and investor in, its fellow UK licensor, Imagination.
The fight follows rather similar lines to that with Intel in CPUs ‘ ARM scores on size and low power/cost, its rival on outright performance. Market analyst Jon Peddie told EETimes: “In general, Imagination has had superior performance to ARM’s Mali, and Mali has had a smaller die size.’ This is why some firms like Samsung and MediaTek use both, depending on the particular requirements of a system-on-chip or device.
The T760 is designed to address that perceived performance shortfall as recent high end Cortex designs have done in the fight against x86. ARM’s graphics product manager, Andy Loats, said: ‘We believe we will far surpass where Imagination is for some time.’ The new design uses two arithmetic pipelines per core, rather than four in the previous high end Mali, the 678, which has not seen significant uptake. That core’s predecessor, the 628, also used two arithmetic pipelines but the 760 delivers a 25% performance boost per core, compared to that offering.
A second Mali was unveiled, this one targeted at midmarket devices and optimized for Android. The T720 promises 533m triangles/second and 4.8Gpixels/second, or 81.6 GFlops in overall performance. It is 30% smaller than previous midmarket Mali cores and uses 15% less dynamic power.
The mobile graphics market is getting more complicated as device differentiation relies more and more on graphical apps and performance. Qualcomm has its own platform, Adreno, acquired from AMD five years ago, as does Intel, though it also uses third party products in some areas (and is an investor in Imagination). Vivante is a small player which has been gaining ground since it switched from the mobile games market to licensing OpenGL-based GPU cores and an even bigger shake-up was promised in June when Nvidia, the king of graphics in the PC space, said it would license its previously closely guarded Kepler GPU technology to other SoC makers.
ARM’s approach is changing as it targets a wider range of segments and as, in all those areas, SoCs become increasingly integrated. This forces it to work very closely with companies who design the elements that coexist with its processors. Earlier this month saw an intensified partnership with DSP core IP supplier CEVA, targeted at devices and communications equipment. And this week it announced a collaboration with ASOCS, a start-up focused on the emerging area of Cloud-RAN.
This indicates the increasingly high levels of integration between different components of a device or base station ‘ modems, processors, GPUs, DSPs and wireless communications are deployed on compact SoCs to reduce space, cost and power consumption. That requires closer cooperation between the providers of chips and IP.
ASOCS licenses its IP (silicon cores for modems) to vendors and chip suppliers in the C-RAN environment and has worked with Intel and China Mobile on the latter’s trials. While it has been vocal in predicting a network architecture that will rely entirely on off-the-shelf servers and IT programming languages ‘ squeezing out specialized DSPs and associated code ‘ it clearly believes that ARM will have a role in those servers alongside the established x86.
Hence the two IP licensors’ joint reference platform, based on ARM Cortex-A57 processors and ASOCS’s Modem Processing Unit, a software defined radio which can be reconfigured for different technologies and standards. The reference design will enable communication chipmakers to develop custom silicon for base stations, a relatively new but important target for ARM, and a segment in which it has won support from Texas Instruments, Freescale, LSI and others.
In particular, the ASOCS alliance looks towards the emerging Cloud-RAN architecture, claiming to offer a significantly lower cost and lower power solution than those used in early stage trials. C-RAN is an important area in which to take a position for ARM, because it is so new and therefore open to new architectures ‘ but so far, early developments have been dominated by Intel, which sees a golden opportunity to expand its data center platforms into the carrier world. But as ARM starts to make a few inroads into the x86 server kingdom, particularly in the cloud sector, C-RAN will be an important bridge between those efforts and its push into base stations. Its CTO is presenting a paper on the ARM Cloud-RAN architecture at this week’s TechCon, the UK firn’s annual developer conference.
ASOCS’ CEO Gilad Garon commented: ‘Solutions based on the ARM/ASOCS reference design will provide silicon designers a highly cost competitive solution while strictly maintaining the full physical layer resources virtualization, while offering the programmability and upgradability of large scale Cloud RAN systems.’
There are two main elements to the start-up’s solution. One is MPL (Modem Processor Language), which allows developers to call C libraries to run jobs on processors and ports DSP-based code into the C-RAN environment. It aims to take a similar approach to that of OpenCL in graphics, shielding developers from the complexities of the hardware below in order to ease migrations to C-RAN.
The other is the MPU (modem processing unit), which provides base station functions on a low cost board. The MPU board is the point of entry for signals from the remote radio head and acts as a bridge to connect to other base stations in the C-RAN architecture, which may be only a few inches away. The approach reduces the cost per LTE sector to under $30 in a large roll-out, says ASOCS.