How efficient is your cruncher?

[[Ion]]

ARM chips are the way of the future for crunching/folding
they offer lower power consumption,they scale better than x86,and the licensing its more liberal than that hot mess with AMD/Intel
and they are cheaper a 4 or 8 core ARM chip on your board of choice is less than 1/4 the cost of a desktop

Maybe. While the tablets do manage pretty good PPD/W, quad-socket systems still are the undisputed winners. Yeah they draw a bunch of power but with remarkably good PPD from it.

 

[Nordic]

I wonder how cheap one of those arm super computer things are. Like 12 quad core arm chips or something.

 

[xvi]

I wonder how cheap one of those arm super computer things are. Like 12 quad core arm chips or something.

One word. Moonshot.

 

[Nordic]

One word. Moonshot.

I said "cheap"

If I could I would take a whole server rack of those for fun.

 

[silentbogo]

I am thinking about building a not-so expensive ARM cluster.
Quad-core RaspberryPi 2 is cheap, but not so fast. I was thinking something in terms of Odroid XU4, but those are in a $75 ballpark and where I'm from the demand for these things is low, so the retail price can be as high as $125.

I only have 2 ARM dev boards left: RPi2 and a Cubietruck. I'll try to set up WCG on both soon and check some results. Would be nice to get one of those 16-core HiSilicon D02's: I bet 64-bit ARMv8 performs much better than older CPUs.

 

[xvi]

Cubietruck

Very interested in the results. Do you know which board specifically you have? I did some digging in to a bit of Rockchip's 8-core ARMv8 RK3368 to compare to my 4-core ARMv7 RK3288 which is good for ~700-800 PPD. The general consensus seems to be that the 8-core is actually slower for some reason in most applications (possibly not mature?).

where I'm from the demand for these things is low

eBay is always the answer here. Parts from China will get there eventually.

 

[silentbogo]

Cubietruck is a relatively old Cubieboard with dual-core ARM A9 (Allwinner A20) / 1GB RAM / 8GB Flash and tons of built-in peripherals (including SATA). Got it awhile ago for experiments, but it came out to be very slow for general use like an office desktop...
I guess I was not paying much attention, while browsing the Hardkernel website, but apparently they have the lowest price on Odroid XU4, but kinda expensive shipping: $75 for a board + $19 international delivery from South Korea. But if I order 10(!!!) I'll only pay $55 for the whole package. With my current monthly salary I'll need to work till next XMass to save an extra $800...

Very interested in the results. Do you know which board specifically you have? I did some digging in to a bit of Rockchip's 8-core ARMv8 RK3368 to compare to my 4-core ARMv7 RK3288 which is good for ~700-800 PPD. The general consensus seems to be that the 8-core is actually slower for some reason in most applications (possibly not mature?).

Maybe it was a circumstantial result. Especially if you are running Android and consider that RK3288 runs at 50%(!) frequency advantage. According to Rockchip their new octa-core CPU runs up to 1.6GHz, which is a 33% performance bump over benchmarked stock (still not that good though).

I just found this comparative benchmark. Looks quite interesting:
http://www.phoronix.com/scan.php?page=article&item=arm_odroidxu_octa&num=1

There is also a nice bench for a pandaboard cluster versus various x86 CPUs. Looks quite interesting, especially in terms of power efficiency.
http://www.phoronix.com/scan.php?page=article&item=phoronix_effimass_cluster&num=1

Tegra K1 (and X1) also looks promising, but it is so expensive, that I'd rather fix-up another used Xeon workstation...

 

[xvi]

Tegra K1 (and X1) also looks promising, but it is so expensive, that I'd rather fix-up another used Xeon workstation...

Yeah. Old Xeon workstations/servers seriously rock the performance/initial-investment category. These ARM and Atom rigs and should win after enough power bills though. Hard to say.

 

[[Ion]]

Yeah. Old Xeon workstations/servers seriously rock the performance/initial-investment category. These ARM and Atom rigs and should win after enough power bills though. Hard to say.

Yeah, the output/W is very good; output/$ not so good. The ODroid UX4 does about 1500 PPD for 9W, for $85. $1000 gets you almost 12 of them (almost) so 18k PPD and 108W. Mind you, $1000 also gets my 64C Opty system, which is, improbably, getting about 60k PPD (or at least would if it was running 24/7, based on points scored / hours of runtime * threads). So, over 3X the PPD, but at nearly 6X the power draw (I suspect it's about 600W at OCed speeds)

 

[silentbogo]

Started my little ARM experiment with crunching on the phone. Now every night I will leave my HTC One with all 4 cores running. I'll try to post results in a few days, once the PPD scores get settled.

I attempted to do the same for RPi2, but forgot that I have no wifi adapter. Fortunately I had some spare CAT-5 and an old D-Link switch, so last night I was battling Windows and its f'n ICS. Bridging networks is functional, but sucks big time in my case: I have at least 6-7 other devices on the net and RPi constantly bumps into an IP conflict.

ICS is a mess, but I'll give it a second try today... otherwise I'll have to find a cheap USB WiFi adapter, or hook it up to my backup router and set it up in repeater mode.
Maybe by the end of the challenge I will have my fleet of devoted ARM minions to do my bidding and take over the world. Moa-ha-ha-ha!

 

[Nordic]

James888 | External| Nvidea Tegra 2 | Integrated | 6w, 100% | 237 | 39.5 | Android 2.4 | Free tablet I picked up

 

[CAPSLOCKSTUCK]

"Wendys Nightmare" is crunching but not a cruncher

X5670 running @ 3.8
CX500 psu
ASUS Rampage ii extreme
12gb corsair triple channel ram
xfx HD 7970

using 252 w ( no monitor)

probably wont crunch consistently throughout the day for long enough to give an accurate ppd but if it does i will tag krusha03 back with accurate figures

 

[xvi]

I'll try to post results in a few days, once the PPD scores get settled.

Update?

 

[silentbogo]

Update?

BOINC client stopped working on my phone 7 days after deployment. Just stuck at logo even after reboot.
So far it was able to do this:
10/27/2015

0:000:05:48:40

198

1
10/26/2015

0:000:09:10:40

201

1
10/24/2015

0:001:19:04:37

901

3
10/23/2015

0:001:22:59:42

1,343

4
10/22/2015

0:000:23:04:40

515

2
10/20/2015

0:000:07:57:27

359

1

The largest result was obtained on a day, when it was plugged in for 24 hours. The rest is a 6-8 hour night charging cycle, so I'd say ~400PPD average (~5W power consumption), but still not enough data to be sure...
On certain days it was skipping tasks altogether, or had a few invalid results, which sucks really bad.

Another downfall is heat. BOINC client on Android is set to suspend all crunching when battery hits 40°C (can be adjusted in settings), which on a hot phone like mine reduces the effective work time to ~50%.

If I fix a motherboard from a THL W200S (Octa-core Mediatek CPU, 1GB RAM), I will try it headless and battery-less to reduce load on cell charger and remove overheating from the equation...

 

[xvi]

Another downfall is heat.

Had the same issue with my phones. I was able to get a little extra performance by taking off the back cover placing an old Thermaltake Orb on my battery (could feel a little heat too!). Helped with the heat throttling a bit.

 

[silentbogo]

Got my phone up and running again for the challenge.

Cubietruck is still resisting due to WiFi issues, but I'm trying to solve it with some spares and soldering iron (my modified WiFi antenna connector got damaged, so no network yet...)
Got it running Android 4.1.1, since there are no alternatives. Haven't messed with configuration yet, but I will also try to attach a big-ass radiator and give it a little boost in CPU and RAM frequencies.

RPi2 is still of no use, but I will try to install Android on it too and figure out the way to run networking through Cubie.

 

[silentbogo]

My army is growing! Neighbor brought a used Galaxy Note, so now I have 10 ARM cores(4+4+2) and 5 Intel cores(10 threads) crunching away. Hopefully this time I'll be able to gather some viable results before it all breaks down.

 

[ThE_MaD_ShOt]

Had the same issue with my phones. I was able to get a little extra performance by taking off the back cover placing an old Thermaltake Orb on my battery (could feel a little heat too!). Helped with the heat throttling a bit.

Freak'n Dragon orb. 7k rpm fan and all. Those things are heavy really heavy. I have one still in the retail packing. Was afraid to use because it is so damn heavy I though it might rip the cpu socket off the board.

Oh on topic, after the 4th rig efficiency went out the window. Don't care no mo.

 

[xvi]

7k rpm fan and all.

Ugh. Seven thousand ear-shredding RPM. I think they did that just so they could sell those Hardcano fan controllers.

Was afraid to use because it is so damn heavy I though it might rip the cpu socket off the board.

I had the same concern with my Zalman CNPS7700-Cu, albeit that was my Sck754 board.

 

[[Ion]]

7000 RPM, hah. That's nothing! The 40MM fans in the 1U system run at 15k+

 

[silentbogo]

7000 RPM, hah. That's nothing! The 40MM fans in the 1U system run at 15k+

My neighbor brought me a managed switch for repair once(I think it was a DES-3200 or something like that).
This bastard had 2 of those! It was making so much noise, I thought I got teleported to the runway of the aircraft carrier.

 

[silentbogo]

So, I have had some disturbances in my crunching schedule, but I've been able to gather some more intel on ARM performance and efficiency.

1. HTC One M7
Tested in 12-16 hour/day sessions, since this is my phone and I have to use it to make calls occasionally.

Snapdragon 600(4 cores, 1.7GHz) | ~700PPD | 5W | 140PPD/W

2. Samsung Galaxy Note 3 Neo Duos
Had this puppy for 3 days. Ran it almost non-stop, but only with 3 cores active.

Snapdragon 400 (4 cores, 1.6GHz) | ~900PPD | 5W | 180PPD/W

3. Cubietruck (CB3)
This one is my lab rat. It went through so much torture, I'm surprised it is still alive. Installed Android 4.2.2, BOINC and SSH server. Ran 24/7 for the past 3 days with 10% overclock, but now it is down for maintenance and relocation.

Allwinner A20 (2 cores, 1.1GHz) | ~255PPD | 3.5W | 72PPD/W

Horrible performance, efficiency is very close to x86 systems, because it has too many power-hungry peripherals. If used with a SATA HDD, then it will consume over 6W and reduce PPW to 42.

Conclusion:
If you have an average family of 4 members and each one of you has an average quad-core smartphone, you can get a combined performance of 2000+ PPD by simply leaving your phones charging overnight (equivalent of a dedicated Celeron G18xx cruncher). If you have tablets, set-top boxes and other insignificant Android-powered devices, you could probably double or triple that number.

The only assumed downfall of crunching on the phone is that it may reduce its battery life due to extended heat exposure. While BOINC suspends computation at 40°C, I'm pretty sure that battery temperature sensors are unreliable and in some devices may not work at all (e.g. generic chinese smartphones, Samsung, Lenovo and majority of ASUS devices). Samsung Galaxy Note was the only device that did not heat up that much. My HTC One was so toasty, I could just flip it upside down and use as a coffee cup warmer.

 

[silentbogo]

Beware, good folk of TPU village! The great necromancer is here!

Just wanted to share some more unimpressive crunchy numbers from my expanded ARMy:
- Just put together a Meizu M2 Note from scraps in my office. Ran BOINC client for some time and got a decent result of 950-1000ppd. It uses a low-end octacore MT6753 chipset, but runs just as good as Snapdragon 415.
- Upgraded my personal phone to Xiaomi Mi4C (also put together from trash for less than $30). It sports a powerful Snapdragon 808, but so far hasn't managed to return any results to WCG.....
If some time during the next week it will finally finish some WUs, I'll post the result here.

P.S. Also got a pair of Nexus7 tablets, but those suck too. The crunching process is always interrupted by either battery temp limit or CPU temp/load, so I might need to turn those into battery-less actively-cooled desktops...

 

[Nordic]

1000 ppd from a phone? That is pretty exceptional ppd/watt, no?

 

[silentbogo]

1000 ppd from a phone? That is pretty exceptional ppd/watt, no?

At max load with fully-charged battery it pulls around 4W from the wall (5V ~0.8A according to my bench PSU), so it's in a ballpark of 240-250ppd/W.
So far that's the best result I've seen from an ARM phone, but I still had no chance to do a proper 24/7 testing on newer Qualcomm chips (Xiaomi Mi4c is my primary phone, so I have to run around with it whole day).
Nexus7 is still resisting. Looked up my BOINC client logs, and apparently it gets interrupted due to CPU/RAM load higher than the set threshold, so I've adjusted those a bit. We'll see if it helps.