Okay, I'm reading through this thread for the first time since it got split or whatever and I'm reading a lot of comments that I feel like remarking on. I'm just going to shoot off a big bullet list of remarks, hope no one minds this.
- The Snapdragon S4 NEON test (
http://www.openssl.org/~appro/Snapdragon-S4.html) needs to specify what NEON operations it's doing, as it is now close to useless. There are NEON ops A15 can't dual issue too. From my limited testing and talks with a couple other people, it looks like A15 can generally do one 128-bit + one 64-bit arithmetic op, and it can do vtbls a lot faster than A8/A9.
- Doubtful that picking one SoC over another is going to make the difference between totally unplayable and fullspeed for Saturn emulators today. New Saturn emulators are needed.
- Before commenting on A8 vs A15 FPU perf (
http://boards.openpandora.org/topic/15560-soc-back-and-forth/page-7#entry312000) I'd like to be certain the A8 version was using NEON, because VFPU is so slow on it (actually, would like to see the assembly in general). I'd also comment that FPU doesn't tend to monopolize mobile app time, including with emulators and even games. For the ones that use it much at all.
- perf/W isn't linear, it's a curve, it also includes static + dynamic elements and varies a ton with implementation. It's totally possible for some Cortex-A15 SoC to give better perf/W than Cortex-A9 at some perf level and worse at some other level. Note that all Cortex-A15s I'm aware of today are on 32nm or 28nm processes, many of which with HKMG, while that's not true for older Cortex-A9 SoCs, so they get that advantage. Comparing peak levels when one can get much higher performance than the other isn't valid, especially considering that some SoCs are applying huge voltage increases to get the last couple 100MHz.
- The SoC vendor, like TI or nVidia, actually has a big influence on the efficiency of a licensed Cortex-A15 outside of the manufacturing process, because they usually implement a soft-macro from ARM (they get it in some kind of RTL), they are responsible for "hardening" the design (implementing the physical layout) and that affects things like perf/W and max clocks. Also, ARM has added power-saving improvements to later revisions of A15 anyway, so the revision used matters (no idea what TI uses)
- The Cortex-M4's are not an alternative to big.LITTLE, nowhere even close. All software could transparently be executed on "little" cores, almost no software will use the M4 cores. The comment that the little cores only save something when the big cores aren't being used isn't really right either - if you use one big core and one little core (as many loads are asynchronous) you'll use less power than 2 big cores (bear in mind both cores have to be clocked at the same speed)
- The Cortex-M4s on OMAP5 do however have an MMU (external to the CPU, unified between the two - also, they have an external cache) and can share address space with normal apps. I bet you anything that this pretty heavily limits their clock speed, although I'm still trying to find out what they can be clocked to. TI doesn't seem to advertise this.
- The comments about various emulators being GPU limited instead of CPU limited needs to be demonstrated with some evidence, especially with Gamecube emulation.
- Qualcomm SoCs don't have big.LITTLE. They do have the ability to clock the CPU cores independently, which ARM-designed CPUs don't. Except you can clock clusters of CPUs independently, ie the A15s vs the A7s in big.LITTLE.
- I'm not part of the Pandora team and I don't really matter compared to notaz. I'll port DraStic whatever is chosen and whatever other emulators I work on in the future, if that's where the concern it. Also, I don't prefer x86, if all other parts of the SoC are equal, that's just not always the case. I'd rather the thing stayed with ARM too if it meant avoiding losing notaz, because there's a much bigger risk that this won't work without him.
Since some people wanted me to post here, I can answer specific questions. If someone has them.
Mentioned A80 in the other thread, ED clearly knows about it now so I guess we'll see if he's interested in getting more information. I do know they should have dev boards available really soon and would probably make one available to him.
with big.Little architecture.
