When Apple announced last September their A7 chip had gone 64-bit, the congregation immediately swooned, but analysts reacted skeptically: “So what? Phones don’t need more memory, and there are no 64-bit apps.” Even pundits miss once in a while, and now the topic is how the chip industry is headed for 64-bit.
Analysts weren’t alone in their initial reaction. Qualcomm marketing leadership completely missed the point too, going so far as a now-former VP saying “there’s zero benefit a consumer gets from [64-bit].” People at large were quick to point out 64-bit Android implementations wouldn’t be ready for some time, but there is varying debate on exactly how much technical work has to be done – the wait may be more in Google organizing a developer rodeo.
This same skepticism applied years ago when ARM introduced a line of 32-bit microcontrollers, at a time when most were 8-bit and a few were 16-bit cores. Who needs that much memory in a small device? I saw that very assertion as recently as yesterday in a discussion forum. My retort: Want your device to be connected? You’d better be looking at 32-bit. Running advanced wireless protocols with encryption algorithms is the order of the day for connected devices.
The push for 64-bit is not, and never really has been, about physical memory for mobile devices. 64-bit can address everything in an NSA data center, clearly a benefit for servers but a stretch for phones and tablets with physical and power consumption limits on memory. Yet, Apple plows in, and the other mobile SoC makers are forced to follow.
64-bit does provide a ton of virtual memory, allowing apps to be essentially kept running and paged in and out of memory as required – say, like in Windows 8, which doesn’t seem to actually kill anything. (Stop laughing, there is a point coming.) It also provides a much wider path through the system, particularly for the GPU, a big boost for apps like computational photography or embedded vision. One big benefit of ARMv8 is it doesn’t choke on 32-bit apps; they run side-by-side with the 64-bit system.
Over the shock and pressing to get something out the door, Qualcomm recently announced their Snapdragon 410 with a quad ARM Cortex-A53 inside, and presumably is burning the midnight oil on a Krait-ish 64-bitter of their own design. Samsung has hinted at but as yet not quite announced Exynos 6, one version featuring Cortex-A53 cores, another with a homegrown Samsung core. Both are likely 2H14 at best for volume availability.
Homegrowing an ARMv8 core seems to be all the rage here, back to that trend I’ve been talking about a lot lately of optimizing processors for workloads. One idea would be 192 GPU cores connected to two ARMv8 64-bit cores, like the upcoming “Denver”-based version of the NVIDIA Tegra K1. Admittedly, there is more posturing here: the first version of Tegra K1 is quad ARM Cortex-A15, with the “Denver” version 2H14 availability.
Putting a desktop-class GPU in a 5W part may be too beefy for most mobile devices, but is probably OK for a larger tablet that can hold a bit more battery. That TDP figure may be a peak number with all 192 GPU cores running full blast; interestingly, according to the folks at AnandTech, the “Denver” version ditches the companion core for some possibly more exotic power-saving techniques. We’ll see when NVIDIA releases silicon into the wild.
Now back to that Windows 8 comment. “Denver” may be the saving point for the Microsoft Surface family, with fresh rumors this week that the Surface 3 will pack the 64-bit Tegra K1 come 4Q14, and ready for Threshold. Unleashing Microsoft on GPGPU technology might finally produce something very, very competitive in the tablet market.
Casting this 64-bit thing as Apple vs. Android is just silly; both will obviously be there, as well as some versions of Windows (8, 9, whatever it takes). Just as 32-bit MCUs swept over the landscape and developers figured out the benefits, 64-bit ARM SoCs will do the same in higher end mobile designs. The question seems to be how ARMv8 plays out – all indications point to more homegrown designs with highly optimized pipeline and interconnect for differentiation, and maybe a return to a dual-core footprint for power and real estate reasons.
lang: en_USShare this post via: