Message boards :
Number crunching :
raspberry pi 3 vs GPU, whats best?
Message board moderation
Previous · 1 · 2
Author | Message |
---|---|
Richard Haselgrove Send message Joined: 4 Jul 99 Posts: 14654 Credit: 200,643,578 RAC: 874 |
When making comparisons of runtime, it's helpful to give an indication of what kind of task you were running at the time - SETI tasks are very variable. Here's a clickable link to your host 7983022: The first result (123 Ksec) is a guppi VLAR: the others are Arecibo mid-AR. |
Kiska Send message Joined: 31 Mar 12 Posts: 302 Credit: 3,067,762 RAC: 0 |
Also while making comparisons is that the wattmeter tells you the consumption per hour. So it would be 0.625 W per hour or 625mW per hour. Therefore 31.25W per task again that is variable though....... |
EEVblog Send message Joined: 20 Apr 16 Posts: 20 Credit: 4,351,842 RAC: 0 |
When making comparisons of runtime, it's helpful to give an indication of what kind of task you were running at the time - SETI tasks are very variable. I'm a SETI noob, so don't understand all this detail yet, sorry. |
EEVblog Send message Joined: 20 Apr 16 Posts: 20 Credit: 4,351,842 RAC: 0 |
Also while making comparisons is that the wattmeter tells you the consumption per hour. So it would be 0.625 W per hour or 625mW per hour. Correct. An average task on the RPi2 takes 50 hours, so 0.625W for 50 hours is 31.25Wh/task. Not including any power supply efficiency loss. |
EEVblog Send message Joined: 20 Apr 16 Posts: 20 Credit: 4,351,842 RAC: 0 |
Although the Rpi solutions might not be the most efficient, if the average time is 50 hours per task, then you only need say 13 boards (4 cores each) to do one task per hour. The Orange Pi One costs $10 for 4 cores, so only $130 in base hardware (not including SD cards, router, PSU, cables etc), and can be passively cooled. But I have not done a bang-per-buck comparison with a PC. Old PC's are often free though, I get plenty in the dumpster, even Core i7's. |
HAL9000 Send message Joined: 11 Sep 99 Posts: 6534 Credit: 196,805,888 RAC: 57 |
Although the Rpi solutions might not be the most efficient, if the average time is 50 hours per task, then you only need say 13 boards (4 cores each) to do one task per hour. The Orange Pi One costs $10 for 4 cores, so only $130 in base hardware (not including SD cards, router, PSU, cables etc), and can be passively cooled. Well for $130 you can get a 750ti GPU that will complete over 100 tasks a day when running 2 tasks at once. SETI@home classic workunits: 93,865 CPU time: 863,447 hours Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[ |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13755 Credit: 208,696,464 RAC: 304 |
[An average task on the RPi2 takes 50 hours, so 0.625W for 50 hours is 31.25Wh/task. Not including any power supply efficiency loss. My GTX 750Ti does (roughly) 2 WU every 25min (roughly). So 12.5min per WU. 60/12.5 gives 4.8 So 4.8 WUs per hour, pulling (roughly) 40W for that hour. 40/4.8 gives 8.3 That's (very roughly) 8.3Wh/task. So a GTX 750Ti is (almost) 4 times as efficient as a RPi2. Grant Darwin NT |
EEVblog Send message Joined: 20 Apr 16 Posts: 20 Credit: 4,351,842 RAC: 0 |
So a GTX 750Ti is (almost) 4 times as efficient as a RPi2. On it's own, yes, but a GPU doesn't work on it's own, it's part of a PC system that also draws extra power. CPU, motherboard, drives, etc. The RPi2 is a complete system on it's own, so it's not an apples to apples comparison. Care to redo the calcs for your complete system? |
jason_gee Send message Joined: 24 Nov 06 Posts: 7489 Credit: 91,093,184 RAC: 0 |
Now that would be interesting, since a 750ti requires a reasonable driving system to feed it to potential. Best experimental GPU codes are on the order of 10% computationally efficient, flops/peak_marketing_flops, (similar with multiple instances and lesser code). Best vectorised CPU code ~20%. The less mature ARM code probably has a small penalty, and the IO subsystem should better suit single multithreaded processes over an instance per core. So I think it could be a pretty fun fight to watch (evenly matched :D) "Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions. |
Kiska Send message Joined: 31 Mar 12 Posts: 302 Credit: 3,067,762 RAC: 0 |
If you want to try that then, my laptop does an mid-AR, in about ~1-2 hour, with a total draw of 30W(kill-a-watt meter), that includes the CPU, mobo, screen(on full brightness), drives and power supply inefficiencies. 30W is with no CPU task running, only the GT840m. 40W with 1 CPU task running |
HAL9000 Send message Joined: 11 Sep 99 Posts: 6534 Credit: 196,805,888 RAC: 57 |
So a GTX 750Ti is (almost) 4 times as efficient as a RPi2. The OPs original question was Raspberry Pi 3 vs GPU. Which is why there is a lot of talk about GPUs. On the GPU side I expected the OP was considering upgrading a GPU in one of their system or adding an addition GPU. On the Raspberry Pi side you can get several for the price of a GPU. However to first the system vs system question. In a previous post I was using the processor TDP values to calculate Watt-hours per task. For two of my systems the results came out to be. For comparison some of my systems To get the complete system power usage I used the power display on my UPS. For my host 5837483 which is a gaming machine. It has several 4 - 3.5" HDDs & a R9 390X. So not really geared toward being super efficient, but does have an 80Plus Platinum PSU. I was reading 97-102w while running 4 CPU tasks. I'll use the high figure of 102w. (102w * 60min)/60)/4 = 25.5Wh per MB task For my host 7324426 which was kind of made with efficiency in mind. It does use a 2.5" HDD, but has an old 350w PSU that isn't even 80Plus certified. I was reading 15-16w while running 4 CPU tasks. (16w * 360min)/60)/4 = 24Wh per MB task Previously I was using 45w for my 750ti GPUs power consumption. Other have observed similar usage form their power meters. So if I were to add the GPU to either one of my two system I referenced the numbers would indeed be different. Since my previous method of ((watts * run time in minutes)/60)/number of concurrent tasks does work when mixing CPU & GPU I'll use daily watt hours/daily number of tasks Host 5837483 base CPU figures (102w*24)/96 tasks a day = 25.5Wh per task Host 7324426 base CPU figures (16w*24)/16 tasks a day = 24Wh per task Host 5837483 with 45w GTX 750ti doing 114 tasks a day added (147w*24)/210 tasks a day = 16.8Wh per task average Host 7324426 with 45w GTX 750ti doing 114 tasks a day added (61w*24)/130 tasks a day = 11.26Wh per task average So if you wanted to build a highly efficient cruncher from scratch you might want to look into getting an ASrock Q1900M & a GTX 750ti. In USD an EVGA GTX 750ti can range from $100-130 depending on which version you get & the MB/CPU runs $70-80 depending on where you shop. Then about another $80-90 for a SSD, 8GB of RAM, & a PSU. So about $300 total. The Raspberry Pi is listed as being $35, but the cheapest I can find it is $47. That does include a power supply, but no SD card. However at $35 8 of them about only be $280. At 50 hours running 4 tasks at once that is about 1.92 tasks per day on average. Scaling that up to 8 devices using 2.5w each would be 20w & 15.36 tasks a day. However that still comes to (20w*24)/15.36 tasks a day = 31.25Wh per task average. Raspberry Pi devices are cheap and consume little power, but are not the most efficient SETI@home devices. So if your goal is to most cost effectively increase your SETI@home contribution they are not the answer at this time. If you already have a Raspberry Pi & want to use it for SET@home there is an application available. SETI@home classic workunits: 93,865 CPU time: 863,447 hours Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[ |
Todderbert Send message Joined: 17 Jun 99 Posts: 221 Credit: 53,153,779 RAC: 0 |
I am enjoying a new Kindle Fire crunching on all four cores @ 1.7w constant. Cost: $40, RAC 263. Not bad for a Pi computing alternative, plus it has a touch screen. http://setiathome.berkeley.edu/show_host_detail.php?hostid=7945727 I own quite a few Raspberry Pis of all makes and models, just have to get around to designing a project for them. |
EEVblog Send message Joined: 20 Apr 16 Posts: 20 Credit: 4,351,842 RAC: 0 |
The Raspberry Pi is listed as being $35, but the cheapest I can find it is $47. That does include a power supply, but no SD card. However at $35 8 of them about only be $280. The Orange Pi One is only $10 + shipping + SD card http://www.orangepi.org/orangepione/ 4 x 1.2GHz ARM cores, slightly more powerful than the Rpi2 |
EEVblog Send message Joined: 20 Apr 16 Posts: 20 Credit: 4,351,842 RAC: 0 |
Yep, seems fairly conclusive. I'm getting 10 x Orange Pi One boards just for kicks and will try a cluster of 40 ARM CPU's and see how it goes. I don't expect the efficiency to be magically greater than the calcs for a single Rpi2 though. |
HAL9000 Send message Joined: 11 Sep 99 Posts: 6534 Credit: 196,805,888 RAC: 57 |
If your plans are to run BOINC on an ARM cluster. You should be aware that BOINC uses shared memory to communicate with apps. So each node will have to run its own instance of BOINC. SETI@home classic workunits: 93,865 CPU time: 863,447 hours Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[ |
©2024 University of California
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.