Last week, we discussed MediaTek’s plans for an octa-core processor that would leverage ARM’s big.LITTLE MP technology and allow up to eight CPU cores to be active at the same time, rather than limiting a smartphone or tablet to four high-speed or four low-speed cores. When asked during Qualcomm’s conference call if his own company might do such a thing, senior vice president Anand Chandrasekher derisively dismissed the possibility, first saying: “You can’t take eight lawnmower engines, put them together and now claim you have an eight-cylinder Ferrari.” When pushed harder, he elaborated. “We don’t do dumb things.”
Ouch. But buried inside that statement are some very interesting assumptions about where the mobile market is headed over the next few years.
The octa-core that isn’t
The first point, of course, is that MediaTek’s upcoming eight core chip isn’t eight lawnmower engines — it’s four Cortex-A15s and four Cortex-A7s that can theoretically run simultaneously if software is optimized to take advantage of that much multi-threading. Remember, the Cortex-A7 is a low-power core that is meant to help conserve power while the Cortex-A15 is for heavy lifting. In theory, these cores can be leveraged in any proportion that delivers the best combination of performance and power efficiency.
MediaTek is betting that it can use a 4+4 CPU combination to deliver optimal performance in both categories. Whether or not that works depends a great deal on low-level software support, but the company isn’t trying to strap eight cheap products together to build a cohesive whole. The Cortex-A15′s power consumption and mobile characteristics may not be as polished as the Snapdragon 800′s, but the chip is far from a pushover in the mobile world. It’s interesting to see a company pushing big.LITTLE MP this quickly after ARM announced it, and it’s even more interesting that the company would be MediaTek, a relatively unknown player.
Hints at Qualcomm’s future strategy
With that said, however, Chandrasekher’s comments underscore the difficulty of ramping multi-threading for mobile software. Remember, the first dual-core phones only shipped three years ago. Nvidia’s Tegra 3 was a quad-core architecture, but the chip underperformed badly in Surface. The majority of smartphones are still dual-core designs, though this is changing a bit at the highest end.
It will take time for big.LITTLE MP to pick up wide software support, and even if that weren’t the case, there’s a real question of where the diminishing marginal return kicks in. The point of big.LITTLE is to save power by scheduling workloads efficiently and moving tasks between the big and little cores as little as possible. Utilizing all cores simultaneously is a necessary component of getting full HSA implementation baked in for future parts, and it undoubtedly helps with boosting performance in certain instances, but the chances of eight-core workloads becoming the norm on tablets are low.
Remember, a core you haven’t powered up at all will always use less power than a core you spin up partway, assuming equal execution time. And this is an extraordinarily sensitive balancing act given that every data copy, every cache flush, every power cycle burns battery power. Furthermore, there’s a very genuine question as to just how many threads a mobile application needs. An application that demands a great deal of background processing is inevitably an application that burns a great deal of battery life, which may mean it’s a poor application, period.
Does that mean 8-core chips are dead? I think not. Remember, H.265 decode will be coming in the future, and it takes considerably more horsepower to handle those tasks than H.264. It’s possible that early octa-core solutions from companies like MediaTek could prove more able than other devices to handle such tasks when software decode becomes available. At the same time, however, that’s a fringe case for a future standard.
I think Qualcomm is right to see octa-core as more of a distraction than a benefit for now, but I also think the benefits being baked into silicon this generation are going to be an important component of maximizing mobile efficiency in years to come. That may not mean eight CPU cores, but it could easily mean eight “real” cores split between CPU and GPU. (Motorola’s so-called 8X doesn’t count.)