3 ms·
How much battery life is saved using an asymmetric multicore design as compared to dynamically downclocking of a "big dog" core? The article does not address th
by cpeterso 15y ago
How much battery life is saved using an asymmetric multicore design as compared to dynamically downclocking of a "big dog" core? The article does not address this, but it does say that smaller die size of the "little dog" core uses 1/3 less power.
- ajross 15y agoIt's a huge "it depends", and on things that aren't ever going to be publicly visible. "Shutting down" a core is itself really complicated. You can drop all the power rails and push the quiescent current all the way to zero, but that loses the cache; so on resume those cache lines have to be refilled, which is work you don't have to do if you maintain power to the cache. Which to choose has to do with how long the CPU will be active for, which depends in sensitive ways on the architecture of the whole system (how much work can be offloaded to the hardware vs. requires interrupt handling in the CPU, etc...). But then of course those other hardware devices have their own power needs. And there's further complexity due to multiple layers of cache, etc... Balancing the whole thing is a big guessing game, and the guesses have to be made by the SoC vendors years before the actual product appears in the market. My intuition, FWIW, is that the asymmetric idea is a non-starter. It sort of exists right now: c.f. the "baseband CPU" (or all the DSP cores you find scattered around these things) in many Android designs, and the extra CPU is really just a huge resource waste. The hardware people will never get this right. Putting all the transistors into a single set of SMP cores gives up on the ability to tune things perfectly, but it puts the OS in control of things. OS vendors are much closer to the problem and better positioned to make good guesses.
- leoc 15y agoThere should be decent public visibility in time, if an A7 ever shows up on a Nexus phone or on a successor to Beagleboard/Pandaboard. Actually, I don't have the link just now, but IIRC someone was trying to write software to do the same thing on the Pandaboard's OMAP4, transferring a running Linux kernel from the A9 core to one of the OMAP4's built-in M3s. (I don't know what they were planning to do for memory protection.)
- caf 15y agoI believe the idea here is that the coherency logic allows the A7 to be presented to the OS as simply another core that happens to run slower and consume less power. This means that the OS is in control - the transition to low-power mode is handled by the OS scheduling processes on the little dogs while putting the big dogs into power saving states.
- sliverstorm 15y agoIn the extreme case of mostly low-activity, switching to a smaller, slower, low-power core is going to save you power. If discrete, the low-power core can be made with completely different parameters and process technology and consume virtually 0 power compared to the "big-dog" in low-power-mode, and you can properly cut power to the leaky "big-dog". However, if you find yourself switching back and forth frequently, you are probably wasting time, power and materials. What is "frequently"? I cannot say.