Glorious CCCG thread -- Cellphone Compute Cluster for GIMPS

[kriesel]

Originally Posted by ewmayer

I'll be interested to see your "baby, it's cold outside" timings, as well.

Likewise. I have some choice weather for such tests too. (Almost wrote "prime weather" ;) Today's high out is comparable to the kitchen's freezer compartment. Bringing in more firewood shortly to augment the silicon space heaters.

[M344587487]

1024K temperature test. Started at ambient ~20C with no airflow until ~14:02, then by an open window ~14C with airflow, then sandwiched between two ice blocks until ~15:30, then back to ~20C ambient no airflow. Attached is a graph of SoC temperature sensors, looks like the ones near the cores were bouncing between 40C and 70C during the ice block challenge.

tl;dr

   20C no airflow primary timings ~0.0524, secondary timings ~0.0582
       14C airflow primary timings ~0.0475, secondary timings  ~0.0482
       ice blocks primary timings ~0.04     secondary timings  ~0.037

Which gives combined synthetic iteration times of ~27.6 ms/it, ~23.9 ms/it and ~19.2 ms/it respectively. What's odd is that the secondary cluster performed much better than the first as temps dropped, to the point where it ended up faster.

0,1,2,3:

INFO: no restart file found...starting run from scratch.
M20000047: using FFT length 1024K = 1048576 8-byte floats.
 this gives an average   19.073531150817871 bits per digit
Using complex FFT radices        16         8        16        16        16
[Jan 25 13:35:29] M20000047 Iter# = 10000 [ 0.05% complete] clocks = 00:08:38.680 [  0.0519 sec/iter] Res64: 9A2AF744DE060296. AvgMaxErr = 0.220879008. MaxErr = 0.312500000.
[Jan 25 13:44:13] M20000047 Iter# = 20000 [ 0.10% complete] clocks = 00:08:44.009 [  0.0524 sec/iter] Res64: D99B4D255F5C0C74. AvgMaxErr = 0.221481918. MaxErr = 0.312500000.
[Jan 25 13:52:58] M20000047 Iter# = 30000 [ 0.15% complete] clocks = 00:08:44.485 [  0.0524 sec/iter] Res64: 284DC13397AAE4DB. AvgMaxErr = 0.221543568. MaxErr = 0.312500000.
[Jan 25 14:01:43] M20000047 Iter# = 40000 [ 0.20% complete] clocks = 00:08:44.671 [  0.0525 sec/iter] Res64: C3ECAE145A41D57D. AvgMaxErr = 0.221724329. MaxErr = 0.296875000.
[Jan 25 14:10:25] M20000047 Iter# = 50000 [ 0.25% complete] clocks = 00:08:42.245 [  0.0522 sec/iter] Res64: 39BDD0E9AB6C5ACB. AvgMaxErr = 0.221664849. MaxErr = 0.312500000.
[Jan 25 14:18:20] M20000047 Iter# = 60000 [ 0.30% complete] clocks = 00:07:54.824 [  0.0475 sec/iter] Res64: A159C329415881C9. AvgMaxErr = 0.221609130. MaxErr = 0.312500000.
[Jan 25 14:26:15] M20000047 Iter# = 70000 [ 0.35% complete] clocks = 00:07:54.622 [  0.0475 sec/iter] Res64: 1204AAA78EC90606. AvgMaxErr = 0.221570026. MaxErr = 0.312500000.
[Jan 25 14:33:59] M20000047 Iter# = 80000 [ 0.40% complete] clocks = 00:07:43.799 [  0.0464 sec/iter] Res64: 2A44C89CB15BEF6C. AvgMaxErr = 0.221480947. MaxErr = 0.312500000.
[Jan 25 14:40:29] M20000047 Iter# = 90000 [ 0.45% complete] clocks = 00:06:30.099 [  0.0390 sec/iter] Res64: 1D7337E424D1D000. AvgMaxErr = 0.221663247. MaxErr = 0.312500000.
[Jan 25 14:47:03] M20000047 Iter# = 100000 [ 0.50% complete] clocks = 00:06:33.219 [  0.0393 sec/iter] Res64: 5EAA63BE1B9FA238. AvgMaxErr = 0.221479401. MaxErr = 0.312500000.
[Jan 25 14:53:34] M20000047 Iter# = 110000 [ 0.55% complete] clocks = 00:06:30.607 [  0.0391 sec/iter] Res64: DA8DD536ABDE47CB. AvgMaxErr = 0.221367128. MaxErr = 0.312500000.
[Jan 25 15:00:17] M20000047 Iter# = 120000 [ 0.60% complete] clocks = 00:06:42.765 [  0.0403 sec/iter] Res64: 642F5B6BC7CEB0D0. AvgMaxErr = 0.221727658. MaxErr = 0.312500000.
[Jan 25 15:07:07] M20000047 Iter# = 130000 [ 0.65% complete] clocks = 00:06:50.504 [  0.0411 sec/iter] Res64: E3553A51D6F6A365. AvgMaxErr = 0.221800594. MaxErr = 0.312500000.
[Jan 25 15:13:54] M20000047 Iter# = 140000 [ 0.70% complete] clocks = 00:06:46.294 [  0.0406 sec/iter] Res64: 56F4FCB4CCCC033B. AvgMaxErr = 0.221403671. MaxErr = 0.312500000.
[Jan 25 15:20:42] M20000047 Iter# = 150000 [ 0.75% complete] clocks = 00:06:48.604 [  0.0409 sec/iter] Res64: 0282DBDDEFF571C8. AvgMaxErr = 0.221367624. MaxErr = 0.296875000.
[Jan 25 15:27:37] M20000047 Iter# = 160000 [ 0.80% complete] clocks = 00:06:54.258 [  0.0414 sec/iter] Res64: 2847475575162412. AvgMaxErr = 0.221594595. MaxErr = 0.312500000.
[Jan 25 15:34:55] M20000047 Iter# = 170000 [ 0.85% complete] clocks = 00:07:18.160 [  0.0438 sec/iter] Res64: 6E883290AB4FE5DD. AvgMaxErr = 0.221736798. MaxErr = 0.312500000.
[Jan 25 15:43:07] M20000047 Iter# = 180000 [ 0.90% complete] clocks = 00:08:11.718 [  0.0492 sec/iter] Res64: D82A2BD113598640. AvgMaxErr = 0.221641509. MaxErr = 0.312500000.
[Jan 25 15:51:48] M20000047 Iter# = 190000 [ 0.95% complete] clocks = 00:08:40.463 [  0.0520 sec/iter] Res64: E13B981CABF386C0. AvgMaxErr = 0.221472874. MaxErr = 0.312500000.
[Jan 25 16:00:32] M20000047 Iter# = 200000 [ 1.00% complete] clocks = 00:08:43.504 [  0.0524 sec/iter] Res64: 99047F27A64C5119. AvgMaxErr = 0.221815586. MaxErr = 0.312500000.
[Jan 25 16:09:16] M20000047 Iter# = 210000 [ 1.05% complete] clocks = 00:08:43.702 [  0.0524 sec/iter] Res64: ABA0B0711FFFACFC. AvgMaxErr = 0.221564638. MaxErr = 0.312500000.
[Jan 25 16:17:59] M20000047 Iter# = 220000 [ 1.10% complete] clocks = 00:08:42.884 [  0.0523 sec/iter] Res64: F3CA77307D4583D2. AvgMaxErr = 0.221703177. MaxErr = 0.312500000.
[Jan 25 16:26:43] M20000047 Iter# = 230000 [ 1.15% complete] clocks = 00:08:43.929 [  0.0524 sec/iter] Res64: BF99A6EF10473797. AvgMaxErr = 0.221476924. MaxErr = 0.312500000.
[Jan 25 16:35:27] M20000047 Iter# = 240000 [ 1.20% complete] clocks = 00:08:43.595 [  0.0524 sec/iter] Res64: 91744165C085E55A. AvgMaxErr = 0.221551273. MaxErr = 0.312500000.
[Jan 25 16:44:12] M20000047 Iter# = 250000 [ 1.25% complete] clocks = 00:08:45.098 [  0.0525 sec/iter] Res64: CCC040796FAF88D2. AvgMaxErr = 0.221474659. MaxErr = 0.312500000.
[Jan 25 16:52:56] M20000047 Iter# = 260000 [ 1.30% complete] clocks = 00:08:43.309 [  0.0523 sec/iter] Res64: 071133B7D2BD4F31. AvgMaxErr = 0.221562684. MaxErr = 0.312500000.

4,5:

INFO: no restart file found...starting run from scratch.
M20000047: using FFT length 1024K = 1048576 8-byte floats.
 this gives an average   19.073531150817871 bits per digit
Using complex FFT radices        32        16        32        32
[Jan 25 13:36:06] M20000047 Iter# = 10000 [ 0.05% complete] clocks = 00:09:16.868 [  0.0557 sec/iter] Res64: 9A2AF744DE060296. AvgMaxErr = 0.247842553. MaxErr = 0.375000000.
[Jan 25 13:45:47] M20000047 Iter# = 20000 [ 0.10% complete] clocks = 00:09:40.720 [  0.0581 sec/iter] Res64: D99B4D255F5C0C74. AvgMaxErr = 0.248834183. MaxErr = 0.375000000.
[Jan 25 13:55:31] M20000047 Iter# = 30000 [ 0.15% complete] clocks = 00:09:43.547 [  0.0584 sec/iter] Res64: 284DC13397AAE4DB. AvgMaxErr = 0.248772266. MaxErr = 0.375000000.
[Jan 25 14:05:11] M20000047 Iter# = 40000 [ 0.20% complete] clocks = 00:09:39.651 [  0.0580 sec/iter] Res64: C3ECAE145A41D57D. AvgMaxErr = 0.248787216. MaxErr = 0.375000000.
[Jan 25 14:14:09] M20000047 Iter# = 50000 [ 0.25% complete] clocks = 00:08:57.896 [  0.0538 sec/iter] Res64: 39BDD0E9AB6C5ACB. AvgMaxErr = 0.248659808. MaxErr = 0.375000000.
[Jan 25 14:22:09] M20000047 Iter# = 60000 [ 0.30% complete] clocks = 00:07:59.679 [  0.0480 sec/iter] Res64: A159C329415881C9. AvgMaxErr = 0.248328552. MaxErr = 0.375000000.
[Jan 25 14:30:11] M20000047 Iter# = 70000 [ 0.35% complete] clocks = 00:08:02.419 [  0.0482 sec/iter] Res64: 1204AAA78EC90606. AvgMaxErr = 0.248587927. MaxErr = 0.343750000.
[Jan 25 14:36:56] M20000047 Iter# = 80000 [ 0.40% complete] clocks = 00:06:44.080 [  0.0404 sec/iter] Res64: 2A44C89CB15BEF6C. AvgMaxErr = 0.248547253. MaxErr = 0.375000000.
[Jan 25 14:42:40] M20000047 Iter# = 90000 [ 0.45% complete] clocks = 00:05:43.846 [  0.0344 sec/iter] Res64: 1D7337E424D1D000. AvgMaxErr = 0.249011792. MaxErr = 0.375000000.
[Jan 25 14:48:23] M20000047 Iter# = 100000 [ 0.50% complete] clocks = 00:05:42.716 [  0.0343 sec/iter] Res64: 5EAA63BE1B9FA238. AvgMaxErr = 0.248647055. MaxErr = 0.375000000.
[Jan 25 14:54:21] M20000047 Iter# = 110000 [ 0.55% complete] clocks = 00:05:57.899 [  0.0358 sec/iter] Res64: DA8DD536ABDE47CB. AvgMaxErr = 0.248535843. MaxErr = 0.375000000.
[Jan 25 15:00:28] M20000047 Iter# = 120000 [ 0.60% complete] clocks = 00:06:07.018 [  0.0367 sec/iter] Res64: 642F5B6BC7CEB0D0. AvgMaxErr = 0.248928904. MaxErr = 0.375000000.
[Jan 25 15:06:42] M20000047 Iter# = 130000 [ 0.65% complete] clocks = 00:06:13.814 [  0.0374 sec/iter] Res64: E3553A51D6F6A365. AvgMaxErr = 0.248515112. MaxErr = 0.375000000.
[Jan 25 15:12:46] M20000047 Iter# = 140000 [ 0.70% complete] clocks = 00:06:04.361 [  0.0364 sec/iter] Res64: 56F4FCB4CCCC033B. AvgMaxErr = 0.248713062. MaxErr = 0.375000000.
[Jan 25 15:19:12] M20000047 Iter# = 150000 [ 0.75% complete] clocks = 00:06:24.898 [  0.0385 sec/iter] Res64: 0282DBDDEFF571C8. AvgMaxErr = 0.249065100. MaxErr = 0.375000000.
[Jan 25 15:25:31] M20000047 Iter# = 160000 [ 0.80% complete] clocks = 00:06:19.361 [  0.0379 sec/iter] Res64: 2847475575162412. AvgMaxErr = 0.248730342. MaxErr = 0.375000000.
M20000047 Roundoff warning on iteration   166998, maxerr =   0.437500000000
[Jan 25 15:31:59] M20000047 Iter# = 170000 [ 0.85% complete] clocks = 00:06:27.822 [  0.0388 sec/iter] Res64: 6E883290AB4FE5DD. AvgMaxErr = 0.248881311. MaxErr = 0.437500000.
[Jan 25 15:40:11] M20000047 Iter# = 180000 [ 0.90% complete] clocks = 00:08:11.603 [  0.0492 sec/iter] Res64: D82A2BD113598640. AvgMaxErr = 0.248942981. MaxErr = 0.343750000.
[Jan 25 15:49:26] M20000047 Iter# = 190000 [ 0.95% complete] clocks = 00:09:14.533 [  0.0555 sec/iter] Res64: E13B981CABF386C0. AvgMaxErr = 0.248550307. MaxErr = 0.375000000.
[Jan 25 15:59:05] M20000047 Iter# = 200000 [ 1.00% complete] clocks = 00:09:39.335 [  0.0579 sec/iter] Res64: 99047F27A64C5119. AvgMaxErr = 0.248845004. MaxErr = 0.375000000.
[Jan 25 16:08:45] M20000047 Iter# = 210000 [ 1.05% complete] clocks = 00:09:39.636 [  0.0580 sec/iter] Res64: ABA0B0711FFFACFC. AvgMaxErr = 0.248549252. MaxErr = 0.375000000.
[Jan 25 16:18:21] M20000047 Iter# = 220000 [ 1.10% complete] clocks = 00:09:35.614 [  0.0576 sec/iter] Res64: F3CA77307D4583D2. AvgMaxErr = 0.248690302. MaxErr = 0.375000000.
[Jan 25 16:27:57] M20000047 Iter# = 230000 [ 1.15% complete] clocks = 00:09:35.461 [  0.0575 sec/iter] Res64: BF99A6EF10473797. AvgMaxErr = 0.248809814. MaxErr = 0.375000000.
[Jan 25 16:37:34] M20000047 Iter# = 240000 [ 1.20% complete] clocks = 00:09:37.210 [  0.0577 sec/iter] Res64: 91744165C085E55A. AvgMaxErr = 0.248753148. MaxErr = 0.375000000.
[Jan 25 16:47:11] M20000047 Iter# = 250000 [ 1.25% complete] clocks = 00:09:36.814 [  0.0577 sec/iter] Res64: CCC040796FAF88D2. AvgMaxErr = 0.248809012. MaxErr = 0.375000000.
  [Jan 25 16:56:51] M20000047 Iter# = 260000 [ 1.30% complete] clocks = 00:09:39.912 [  0.0580 sec/iter] Res64: 071133B7D2BD4F31. AvgMaxErr = 0.248568909. MaxErr = 0.375000000.

I wouldn't recommend running too cold, I tried the test again at 2560K but I don't think the SoC ramped up to compensate as the phone errored out and stopped execution saying it was too cold for the battery. When it warmed up it started executing again.

Originally Posted by ewmayer

Are you able to edit /proc/cpuinfo? I don't know how the OS exposes/hides cores from Joe User, but it might be worth a shot to see if simply exposing them via that system file allows a user app to use them. I would need to send you a little hack of the Mlucas util.c file to test this - one of the pending changes in v18 is to work around portability issues by having the program read the relevant data directly from that file rather than relying on the ABI for it. I would be surprised if said hack actually worked in terms of giving you access to cores 6:7, but it's easy enough to try.

That would be a hilarious solution if it worked, unfortunately not. It let me edit and save the file but it was instantly overwritten. It's a virtual file created by the kernel to advertise the hardware, I'm guessing that even if the edit stuck it would only trick userland programs that used cpuinfo not the kernel itself. A custom kernel is required, hopefully it exists.


BTW 1024k powered unlocked -cpu 0:5 failed to produce any results, it was busy but I think a process cannot be on both clusters at the same time so it was probably busy juggling threads (over half an hour with no results, when there should have been at least 3). There appear to be a few different ways big.little works which might explain it: https://community.arm.com/processors...out-big-little

[ewmayer]

M344587487: Those are some very nice speedups due to cooling - so the obvious 'production' issue is to open the case so as to be able to stick mini-cooling-fin squares atop just the CPUs and get some airflow running parallel to the cooling channels in the fins.

By way of reference, my Odroid C2 needs ~40 msec/iter @1024K and 120 msec/iter @2560K, it's doing DCs at the latter runlength. Your timings-with-cooling extrapolate to only little less total throughput across both CPUs. Not bad, methinks.

[M344587487]

Adding a direct heatsink looks to be a pain at least for the S7 Edge: https://www.ifixit.com/Teardown/Sams...Teardown/56845

You open the phone from the back, but the SoC is on the wrong side of the motherboard facing the screen. To put a heatsink directly on the SoC you need to keep everything splayed out if the connections between boards even allows for that.

Originally Posted by masser

I may have an iPhone 5S to donate to the cause. Is it too old? Would any one be able to provide a link to instructions for wiping all personal data from a phone?

Wikipedia says it has the first 64 bit ARM chip used in phones, but it's only a 1.3GHz quad core on a 28nm process, probably raspberry pi 3 levels of performance. It's in the grey area of technically being usable if it can be bent to our will but not really being worth bothering with IMO, YMMV.

[ewmayer]

Originally Posted by M344587487

Adding a direct heatsink looks to be a pain at least for the S7 Edge: https://www.ifixit.com/Teardown/Sams...Teardown/56845

You open the phone from the back, but the SoC is on the wrong side of the motherboard facing the screen. To put a heatsink directly on the SoC you need to keep everything splayed out if the connections between boards even allows for that.

Is there any kind of thermal heat-spreading between the 'wrong side' - or between either side - of the CPUs and the metal parts of the case? If not, that means that your cooling efforts yesterday were all indirect through-the-case. The fact that such gave good speedups suggests that simply opening the back of the case and getting air flowing over the guts of the phone might give similar timings to the best ones you got yesterday.

Oh, also I misspoke re. the Odroid C2 comparison - if your best timings @1024k are ~40 msec/iter for each of the 2 per-CPU jobs that is in fact roughly 2x the throughput of the C2, which is pretty good.

[M344587487]

There's a heatpipe connected to the metal frame, looks like a decent attempt at cooling: https://d3nevzfk7ii3be.cloudfront.ne...JN3FQhjVl.huge


I won't dismantle this one to investigate further as it'll probably kill the screen.

[kriesel]

Originally Posted by M344587487

There's a heatpipe connected to the metal frame, looks like a decent attempt at cooling: https://d3nevzfk7ii3be.cloudfront.ne...JN3FQhjVl.huge

I won't dismantle this one to investigate further as it'll probably kill the screen.

Bonding a heat sink to the outside of the case other side of the case wall from the heat pipe may help. It's better than mere natural convection off a limited surface area. A little cpu cooler fan, on the heat sink, even more so, though that ups cost and power a bit.

[ewmayer]

Originally Posted by M344587487

There's a heatpipe connected to the metal frame, looks like a decent attempt at cooling: https://d3nevzfk7ii3be.cloudfront.ne...JN3FQhjVl.huge

I won't dismantle this one to investigate further as it'll probably kill the screen.

It stands to reason that getting some airflow over the heatpipe or the surface it spreads the heat to should help - can you also post a pic of the other half of the phone, the one that mates with the half-clamshell shown in yor pic? I'm trying to get a sense of the 3D structure of things. Best would be a shot laid out the same way. Thanks!

[M344587487]

It's all detailed in the ifixit teardown link a few posts up, here's the side of the motherboard that sits on the copper pipe. The red square is the SoC with RAM above it: https://d3nevzfk7ii3be.cloudfront.ne...QQXRVyoHT.huge

[ewmayer]

Originally Posted by M344587487

It's all detailed in the ifixit teardown link a few posts up, here's the side of the motherboard that sits on the copper pipe. The red square is the SoC with RAM above it: https://d3nevzfk7ii3be.cloudfront.ne...QQXRVyoHT.huge

Thanks - so if I'm mentally fitting that together properly, that row of various subsystem-processors including the SoC sits atop the heatpipe, whihc distributes the heat to metal back of the case?

[xilman]

Originally Posted by M344587487

1024K temperature

:woot: That's bloody hot!