SMT Solving on an iPhone (opens in new tab)

> The i7-7700K is a desktop CPU drawing at least 45 watts of power and clocked at 4.5 GHz when running a single-threaded workload; in contrast, the iPhone was unplugged, probably doesn’t draw 10% of that power, and runs (we believe) somewhere in the 2 GHz range

A12 runs at 2.5Ghz, and draws ~3.64W of power. I also think the actual power draw of the i7-7700K may be north of 45W, Anandtech claims 90W (with a TDP of 91W).

I bet if you could pop a heatsink on the A12(X) and overclock it/up the TDP, it would come very close or beat intel at most single-threaded applications.

Source: https://www.anandtech.com/show/13392/the-iphone-xs-xs-max-re...

a. single threaded code

b. extended benchmark duration - no thermal throttling encountered

c. straight forward C compiler - no special code for the GPU unit or whatever

d. main reason is probably memory and cache configuration

I think it is time to consider A(RM) and Intel cores at par. Qualcomm cores maybe X% slower but they will close the gap.

It is now a question of system trade-offs, as designers jangle memory sizes, cache sizes, cores, cooling, vector extensions, etc

This is a really pivotal event.

And dont forget AMD challenging Intel on like-for-like cores.

Its pretty amazing that noone has put together a real ARM alternative for the server side.

Thanks for this - with the release of the A12X Bionic, I've wanted to do something similar - i.e. build some specialized tools to leverage that CPU for atypical / non-mobile use cases.

I wish Apple would get on with it and include Terminal.App and the FreeBSD user land and Xcode for iOS so we can start running a shell and doing interesting things with this new hardware :-)

I'd definitely believe that cache plays a big role -- while I don't know nearly as much as the author about how SMT solvers work, I've seen how much Z3 bogs down when two instances are running in parallel. That's quite consistent with the cache explanation.

I really like the new Apple hardware ; it's fast enough to do most things on (like this example), but there are no real serious production apps for developers. I have been building my own tooling for iPad but it's really not something that can be released. I do use the iPad as daily driver for most things, but writing code on it is still not very feasible. There is no way (as far as I know) to 'catch' actual crashing processes, for instance if you make a boo boo when coding something with UIKit (check Continuous editor for iPad for an example of something that will be crashing completely all the time when/if you try anything serious on it), the entire app will crash and that's not really workable. If it is possible to actual prevent that, it would be possible to make an actual interpreter with a debugger and make a programming environment on it.

But for experiments and in-house tooling, it works very well, and I barely have to touch my laptop ever because I have enough tooling to do most things I need to do.

Small things are; more developer freedom (allowing fuller control; probably will happen very very slowly) and better keyboard (for the iPad Pro).

Exciting. And the cache size is likely the main reason. But that can not be concluded from the analysis.

>>VTune agrees, and says that Z3 spends a lot of time waiting on memory while iterating through watched literals.

But how much is "a lot" is not specified. And the propagation also allocates memory sometimes, to keep the learned clauses. Not sure how Z3 manages this, but couldn't it be that the mallocs are just slower on that desktop, so an OS issue?

It's worth keeping in mind that Intel tied architecture to process node, and have been stuck on 14nm for many years. Very little changes have been made to Skylake (2015) by Intel even with the latest 9th gen release. They're starting to look long in the tooth on many fronts, AMD's SMT is more efficient, and AMD has a much better boost algorithm as well (Precision Boost 2), among other aspects.

Had their roadmap not tied node to arch, this comparison wouldn't look so damning. Also, many engineers left Intel for Apple because they're talented and tired of releasing Skylake every year. Once Intel fully decouples node and arch, we should see non-mobile CPUs running away from mobile chips like the A12.

This benchmark is essentially Intel's 2015 design on 14nm vs Apple's 2018 design on 7nm.

Could Apple's next "innovation" not be a consumer product but to own the chip world? The iPhone "just" being the beta test for the chips?

I think this probably has a little to do with AMD. AMD wasn't there to push intel and they introduced incremental upgrades for a while. I remember when athlon64 was introduced it really kicked intel into gear. At this point it might be that there might be less you can do to improve performance since we might be hitting semiconductor limits, but I think a lack of competition probably didn't help/