Processor design is all about mixing and matching different cores and accelerators to achieve the perfect balance of performance, power efficiency, graphics and functionality. This is a complex task which AMD hopes to simplify with its HSA (Heterogeneous Systems Architecture), and in so doing, greatly enhance its position in the SoC (system-on-chip) space from smartphones to supercomputers. However, it is just establishing HSA as major SoC vendors are developing comparable capabilities using their own architectures, which could leave the x86 challenger having to battle for air alongside Intel, Nvidia and Apple.
While SoC makers increasingly integrate multiple CPU and graphics cores, these are hard to program because they have separate memory pools. The problem can worsen when other other accelerators for specialized tasks, such as DSPs (digital signal processors), are added too. HSA enables CPU and GPU cores ‘ and other elements in future ‘ to share resources such as memory. AMD has attracted strong industry support for its platform, and will ship the first commercial HSA product itself in January.
Although AMD has often presented HSA in the context of mainstream computing and supercomputing, there is intense interest from the mobile sector, where devices are packing an increasingly wide variety of functions and chips, and where a highly integrated, power efficient SoC is key to success. ARM, Imagination, Mediatek, Qualcomm and Samsung are among the members of the HSA project.
Imagination plans to support the architecture on all its cores, though it has not given a timescale, but this could aid its efforts to expand from being a GPU core specialist to ‘ via the acquisition of MIPS ‘ providing mixed-core platforms. ARM’s CTO Mike Muller, speaking at a recent AMD developer conference attended by EETimes, was more cautious, saying: “HSA is good for midrange systems, but I’m not sure if it scales to the smallest systems, and I don’t think it does scale to high performance computing with tens of thousands of processors.’
As well as SoC vendor backing, HSA would benefit greatly from native operating system in iOS, Android, Windows and Chrome, though the OS giants have not yet made any commitments. Indeed, there may be conflicts for Apple, which increasingly wants to control its own chip platforms, and which is a major backer of the OpenCL specifications. These focus on programming in a standard way on parallel arrays of GPUs, but the body behind them, Khronos Group, is starting to include some of the same capabilities as HSA, including shared memory.
Intel and Nvidia also back OpenCL, and the latter told EETimes that the platform, as well as Nvidia’s own CUDA software environment, will increasingly offer features that mirror HSA’s. CUDA has mainly been implemented in the supercomputing environment, but Nvidia promises to adapt it for its Tegra mobile SoC to support high end graphics and mixed-core systems. The latest versions of CUDA already support some HSA-like flat memory features.
Phil Rogers, president of the HSA Foundation and an AMD fellow, says the advantage of the AMD design will lie in its flexibility. Rather than adapting a platform designed for graphics, it was conceived with many types of chip design in mind. “We started first with GPUs, but other accelerators are expected, including DSPs, DMA engines, codecs and crypto processors,’ he told the conference. ‘Each company will migrate to accelerators as it makes sense.”
The overall effort, whichever firms are leading them, will lead to simplified development of multicore SoCs and the applications to harness them, and that will help make sophisticated mobile functionality more affordable and power efficient. But for AMD it is essential that its HSA takes the lead in this, or it risks being irrelevant in the mobile space. It largely missed the mobile boat but did have a vision of mixed-core SoCs before anyone else, when it bought GPU maker ATI Technologies in 2006 and went on to develop an integrated CPU/GPU offering for PCs.
There is little hope of it winning over Intel and Nvidia to its side, so it needs to stay a step ahead in functionality and developer support. And of course, Qualcomm, ARM and Samsung are powerful friends to have in an increasingly mobile-driven world, while Nvidia’s future position is uncertain still. It is making the right moves, leveraging its high end graphics expertise for advanced mobile platforms and integrating the modem technology it acquired with Icera ‘ also cutting edge ‘ with its processors to go after Qualcomm.
The results so far have been up-and-down, but that integrated processor/modem will help it move out of its stronghold in tablets into lower cost, lower power mobile devices. The first products to run on the all-in-one Tegra 4i will ship in the first half of next year, CEO Jen-Hsun Huang said at the company’s recent quarterly results announcement, and that should have a marked impact on its mobile performance. The 4i was recently certified by AT&T, which is important, according to Huang, who has said the US carrier is ‘the most rigorous and if we get through AT&T, we’re pretty much set on just about all the western world’s open markets’.
Despite the AT&T kitemark, the 4i devices will initially be targeted outside the US, in areas where there is no CDMA to complicate the modem picture. A competitive advantage for Tegra 4i will be the software-defined nature of the Icera modem technology, which allows operators to reconfigure devices for new frequencies even after purchase.
Some analysts are unconvinced by the Tegra strategy though. Wang and Needham analyst Rajvindra Gill told Reuters that the firm should sell or spin off Tegra and stick to higher margin markets such as gaming PCs, automotive and enterprise computing. “Become a much higher margin company and continue to return capital to shareholders,’ he urged.
The third quarter results highlighted the mixed feelings around the Tegra mobile processor SoC. Overall company revenue was up by 7.9% compared to the second quarter, reaching $1.054bn, and this figure was far outstripped by Tegra, which was up 111.4% to $111.2m, thanks to the shipping of Tegra 4. However, year-on-year, revenues were down 12.5%, and those from Tegra fell by a huge 54.4%. Nvidia said the annual decrease was down to lower volume shipments of Tegra 4 compared to its predecessor Tegra 3, while chips for gaming PCs and other high end applications remained the mainstay of revenues.
The sequential uptick reflected the shipping of Tegra 4 as well as increased sales in embedded and automotive sectors. Tegra may hit the headlines in tablets but 25% of its revenues come from the car industry and this was the only sector to see a year-on-year revenue rise. Nvidia expects this trend to intensify. “From digital clusters to infotainment systems to, in the future, driver assistance, the car is going to be an opportunity for more than just one GPU,” said Huang on the analyst conference call.