Glorious CCCG thread -- Cellphone Compute Cluster for GIMPS - Page 5

M344587487 · 2019-01-27, 20:04

As you look from the front vertically the SoC is mid-right of the screen. There's the screen including LCD, then a metal frame with a copper heatpipe spanning the chips on the motherboard including the SoC facing you, the back of the motherboard has some metal that sinks heat into the battery cover, with the metal back behind that and finally the glass back of the phone.

xilman · 2019-01-27, 20:04

Quote:

Originally Posted by M344587487

1024K temperature

That's bloody hot!

PhilF · 2019-01-27, 22:07

Quote:

Originally Posted by xilman

That's bloody hot!

I totally missed that. I hope not too many people were killed.

retina · 2019-01-28, 04:24

I suggest that if you use these things long term 24/7 at raised temperatures to remove the battery. And if you still need to cover power outages then connect the PSU to a UPS. Small sized UPSes are easy to find.

M344587487 · 2019-01-29, 22:41

Got two S7 Edge's rooted, working and benched. Setup a standardised test bench of a wind tunnel with a 12cm PC fan one end with the phone standing on end in the middle of the tunnel with the back facing the fan. Simulates best case "cluster in a PC case" setup, albeit with an unopened phone. This could estimate a less favourable airflow with an opened/heatsinked phone. The plan is that if and when I get other phones working I'll be able to compare SoCs with as fair conditions as is feasible. Encountered an S6 with a locked bootloader which is unrootable as a result, something else to look out for particularly on S6's and similar era.

Synthetic combined ms/it:

Code:

S7 Edge Device A
 1024K  23.2533763576263
 2560K  50.526165810006
 4608K  88.9266271051403
 7680K 150.192587255844
18432K 404.475944826566

Code:

S7 Edge Device B
 1024K  21.2311242522294
 2560K  52.2090463973316
 4608K  89.0637352324169
 7680K 150.247721728181
 18432K 401.203267403529

Similar results which is nice. These synthetic timings took the average of 10K iterations, ignoring the first 10K, running simultaneous FFTs on each cluster. The tests were performed for 1, 2, 3, 4 and 10 hours respectively.

I'm dumping /proc/cpuinfo and /proc/version, what else is useful to dump?

Mark Rose · 2019-01-29, 22:51

Do you have any idea how many watts they're drawing?

M344587487 · 2019-01-30, 12:26

I've made a few scripts to help anyone that wants to convert a big.LITTLE device into an mlucas compute node. Only tested on two S7 Edge's so far: https://github.com/sillygitter/c3tools

The setup script takes a user defined description of the clusters in the format used by mlucas's -cpu flag (the description for each cluster is passed verbatim to mlucas), creates the worker directories, generates the mlucas.cfg for each cluster with the other clusters under load, and modifies the timings in mlucas.cfg to be contiguous as they're unreliable as generated due to potential aggressive governor behaviour on big.LITTLE devices.
The bench script performs a 20 hour simultaneous FFT benchmark test with 1024K (1 hour), 2560K (2 hours), 4608K (3 hours), 7680K (4 hours) and 18432K (10 hours) FFT lengths, putting the results and a few vitals from the device into a tarball.

Quote:

Originally Posted by Mark Rose

Do you have any idea how many watts they're drawing?

Under the same conditions as the above test (wind tunnel, screen permanently on at minimal brightness, fully charged battery) with simultaneous 1024K runs, device B is bouncing between 4.5W and 8.0W. Not including the PC fan which is powered by a PC PSU. The wattmeter has a resolution of 0.5W, showing 0.0W with nothing plugged in and 0.5W with the 5V/12A USB hub plugged in with no devices attached. The simultaneous 1024K test is running for a while so that an average can be generated however the KWh resolution is 0.01, it'll take a while to tick over and may take longer for an accurate result.

Quote:

Originally Posted by retina

I suggest that if you use these things long term 24/7 at raised temperatures to remove the battery. And if you still need to cover power outages then connect the PSU to a UPS. Small sized UPSes are easy to find.

Good point. These devices are not meant to have the battery removed so I don't know if they remain powered without one. Failing that there are apps that limit how charged the battery can get.

Quote:

Originally Posted by xilman

That's bloody hot!

It's ok I'm using oven gloves ;)

M344587487 · 2019-01-30, 18:49

Quote:

Originally Posted by Mark Rose

Do you have any idea how many watts they're drawing?

Ran the simultaneous 1024K test for a while, it's taking ~118 minutes to consume 0.01 kWh, giving a power consumption of ~5.085W unless my maths is off. It's still running so the figure may get slightly more accurate, but it's ticked over three times now and they've all been ~118 minutes for 0.01 kWh.

ewmayer · 2019-01-30, 19:38

Very nice - thanks for the data and the utility scripts! The timings you posted yesterday for 1024-18432k FFT indicate quite good scaling - the 7680/1024 timing ratio is just over 7, slightly better than the 7.5 one gets from simple linear extrapolation and appreciably better than the 8.6 one expects based on an n log n extrapolation. On my Odroid C2 the same FFT lengths yield a timing ratio of 11, i.e. there is significant deterioration-versus-theoretical at the larger FFT lengths.

Re. cooling, your case disassembly and various experiments indicate that typical phones like these actually have good heat transfer from the CPUs and other chips to the case, the main issue is getting decent airflow over the case. That is a favorable result, since it means one can avoid hardware hackery involving attachment of heat sinks, which is ruled out on many (most?) phones due to the flipped chips inside. Stick ~10 phones into some kind of simple mini-rack with between-phone spacing roughly equal to phone thickness, use a 125mm case fan to blow air through the stack.

So it seems the main remaining question is in regard to the "target device demographic", i.e. the acceptable kinds of used-phone hardware and where one might be able to obtain such in volume at an acceptable price. Due to the internal-heat-transfer and USB-debug-mode issues we need the cases to be reasonably intact and the screens perhaps cracked, even badly, but still usable for simple input such as needed to enable USB debug mode for the needed I/O. Battery can be dead/missing. So perhaps phones a couple generations older than bleeding edge, the kind users ditch when the screen gets damaged, the battery stops taking a charge, or simply when they want a more recent model?

M344587487 · 2019-02-01, 13:34

Quote:

Originally Posted by ewmayer

...
So it seems the main remaining question is in regard to the "target device demographic", i.e. the acceptable kinds of used-phone hardware and where one might be able to obtain such in volume at an acceptable price. Due to the internal-heat-transfer and USB-debug-mode issues we need the cases to be reasonably intact and the screens perhaps cracked, even badly, but still usable for simple input such as needed to enable USB debug mode for the needed I/O. Battery can be dead/missing. So perhaps phones a couple generations older than bleeding edge, the kind users ditch when the screen gets damaged, the battery stops taking a charge, or simply when they want a more recent model?

Been scouring ebay for a few weeks now and have bought two S7 Edge ~£26 each, two S6 ~£20 each and one A8 ~£12, they tend to go for more than that I'm just a patient cheapskate (it helps that compute is not the typical use case so we have different criteria to most buyers). Unfortunately the A8 and one S6 can't get into android due to FRP, the other S6 has a locked bootloader so cannot be rooted (as does another S6 looked at, S6's may not be suitable full stop if most have locked bootloaders, I recall hearing that many phones of a certain era had locked bootloaders and Exynos S6's may be part of that). That makes the blind success rate 40% but now that some of the common issues are known that should improve, bugging sellers with questions is a pain but necessary.

I think the majority of the reason the timings and power draw are decent for the S7 is the relatively new process node. I'm looking for 20nm at worst but 28nm should be tested to confirm, when you get down to 28nm some of the models are ARMv7 not v8 so you need to be extra careful. There are also a lot of quad cores at 28nm to avoid, not enough compute and too much overhead IMO. There are some phones that have three clusters, notably phones using the Helio X20 SoC or better (https://en.wikichip.org/wiki/mediatek/helio ). Three clusters may change the equation somewhat, I've ordered a phone that should contain one but another issue is that some models use multiple SoCs depending on availability at the time I guess. It may be that devices with three clusters cannot have all clusters active at once which would be a shame.

ewmayer · 2019-02-01, 20:16

Poking around on ebay US using search string "samsung s7 edge for parts", here a couple typical-seeming hits:

1. https://www.ebay.com/itm/Samsung-Gal...Rvb:rk:10:pf:0 -- But It Now only @$70:

"LCD is BAD
The Phone was originally for Verizon, and is Unlocked for any GSM Carrier
The battery door / back is missing
The Sim Card tray is missing
The Touch Screen still works, as I am able to push buttons and navigate through the phone"

2. https://www.ebay.com/itm/Samsung-Gal...ThGI:rk:9:pf:0 -- Current bid $6.50, but meaningless since 2 days left. Buy It Now price set at $99:

"Samsung Galaxy S7 edge SM-G935X - 32 GB - Gold Platinum Smartphone. Condition is For parts or not working. Shipped with USPS Priority Mail Padded Flat Rate Envelope.

Device was a demo unit and has heavy damage. Screen is broken and has screen burn from being on all day. As is, no returns."

Those seem rather pricy vs what you paid, so "patient cheapskate" is the correct m.o. I'm hopeful that with time and more people working the problem we'll identify a few more promising "veins of ore" - given how recent this thread is, we are still very much in the "proof of principle" stage.

In one sense - unrelated to GIMPS work - I'm actually glad of the higher-than-expected prices for damaged gear, as it means there is a lot of reburbishment and secondary-usage going on, i.e. you don't simply have still-good-for-parts gear going straight to the landfill. The longer the old gear gets used, the less new gear needs to be manufactured, with all the environmental impacts attendant to that.

2019-01-28, 04:24	#48
retina Undefined "The unspeakable one" Jun 2006 My evil lair 2⁶·3·5·7 Posts	Don't cook the battery I suggest that if you use these things long term 24/7 at raised temperatures to remove the battery. And if you still need to cover power outages then connect the PSU to a UPS. Small sized UPSes are easy to find.

2019-01-29, 22:41	#49
M344587487 "Composite as Heck" Oct 2017 1652₈ Posts	Got two S7 Edge's rooted, working and benched. Setup a standardised test bench of a wind tunnel with a 12cm PC fan one end with the phone standing on end in the middle of the tunnel with the back facing the fan. Simulates best case "cluster in a PC case" setup, albeit with an unopened phone. This could estimate a less favourable airflow with an opened/heatsinked phone. The plan is that if and when I get other phones working I'll be able to compare SoCs with as fair conditions as is feasible. Encountered an S6 with a locked bootloader which is unrootable as a result, something else to look out for particularly on S6's and similar era. Synthetic combined ms/it: Code: S7 Edge Device A 1024K 23.2533763576263 2560K 50.526165810006 4608K 88.9266271051403 7680K 150.192587255844 18432K 404.475944826566 Code: S7 Edge Device B 1024K 21.2311242522294 2560K 52.2090463973316 4608K 89.0637352324169 7680K 150.247721728181 18432K 401.203267403529 Similar results which is nice. These synthetic timings took the average of 10K iterations, ignoring the first 10K, running simultaneous FFTs on each cluster. The tests were performed for 1, 2, 3, 4 and 10 hours respectively. I'm dumping /proc/cpuinfo and /proc/version, what else is useful to dump?

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Official GIMPS 2019 fundraiser thread	Prime95	GIMPS Board of Directors News	46	2019-11-22 06:01
GIMPS wiki account request thread	ixfd64	mersennewiki	169	2018-09-21 05:43
Cuda and a cluster	efiGeek	Msieve	17	2015-12-06 14:31
GPUs vs. Cellphone GSM	Rodrigo	GPU Computing	2	2011-07-14 07:48
Cluster @ MSRC	smh	NFSNET Discussion	1	2003-08-12 08:52

2019-01-27, 20:04	#45
M344587487 "Composite as Heck" Oct 2017 1110101010₂ Posts	As you look from the front vertically the SoC is mid-right of the screen. There's the screen including LCD, then a metal frame with a copper heatpipe spanning the chips on the motherboard including the SoC facing you, the back of the motherboard has some metal that sinks heat into the battery cover, with the metal back behind that and finally the glass back of the phone.

2019-01-29, 22:51	#50
Mark Rose "/X\(‘-‘)/X\" Jan 2013 C52₁₆ Posts	Do you have any idea how many watts they're drawing?

2019-01-30, 19:38	#53
ewmayer ∂²ω=0 Sep 2002 República de California 5·2,351 Posts	Very nice - thanks for the data and the utility scripts! The timings you posted yesterday for 1024-18432k FFT indicate quite good scaling - the 7680/1024 timing ratio is just over 7, slightly better than the 7.5 one gets from simple linear extrapolation and appreciably better than the 8.6 one expects based on an n log n extrapolation. On my Odroid C2 the same FFT lengths yield a timing ratio of 11, i.e. there is significant deterioration-versus-theoretical at the larger FFT lengths. Re. cooling, your case disassembly and various experiments indicate that typical phones like these actually have good heat transfer from the CPUs and other chips to the case, the main issue is getting decent airflow over the case. That is a favorable result, since it means one can avoid hardware hackery involving attachment of heat sinks, which is ruled out on many (most?) phones due to the flipped chips inside. Stick ~10 phones into some kind of simple mini-rack with between-phone spacing roughly equal to phone thickness, use a 125mm case fan to blow air through the stack. So it seems the main remaining question is in regard to the "target device demographic", i.e. the acceptable kinds of used-phone hardware and where one might be able to obtain such in volume at an acceptable price. Due to the internal-heat-transfer and USB-debug-mode issues we need the cases to be reasonably intact and the screens perhaps cracked, even badly, but still usable for simple input such as needed to enable USB debug mode for the needed I/O. Battery can be dead/missing. So perhaps phones a couple generations older than bleeding edge, the kind users ditch when the screen gets damaged, the battery stops taking a charge, or simply when they want a more recent model?

2019-02-01, 20:16	#55
ewmayer ∂²ω=0 Sep 2002 República de California 5×2,351 Posts	Poking around on ebay US using search string "samsung s7 edge for parts", here a couple typical-seeming hits: 1. https://www.ebay.com/itm/Samsung-Gal...Rvb:rk:10:pf:0 -- But It Now only @$70: "LCD is BAD The Phone was originally for Verizon, and is Unlocked for any GSM Carrier The battery door / back is missing The Sim Card tray is missing The Touch Screen still works, as I am able to push buttons and navigate through the phone" 2. https://www.ebay.com/itm/Samsung-Gal...ThGI:rk:9:pf:0 -- Current bid $6.50, but meaningless since 2 days left. Buy It Now price set at $99: "Samsung Galaxy S7 edge SM-G935X - 32 GB - Gold Platinum Smartphone. Condition is For parts or not working. Shipped with USPS Priority Mail Padded Flat Rate Envelope. Device was a demo unit and has heavy damage. Screen is broken and has screen burn from being on all day. As is, no returns." Those seem rather pricy vs what you paid, so "patient cheapskate" is the correct m.o. I'm hopeful that with time and more people working the problem we'll identify a few more promising "veins of ore" - given how recent this thread is, we are still very much in the "proof of principle" stage. In one sense - unrelated to GIMPS work - I'm actually glad of the higher-than-expected prices for damaged gear, as it means there is a lot of reburbishment and secondary-usage going on, i.e. you don't simply have still-good-for-parts gear going straight to the landfill. The longer the old gear gets used, the less new gear needs to be manufactured, with all the environmental impacts attendant to that.