| Index | Recent Threads | Unanswered Threads | Who's Active | Guidelines | Search |
 |
World Community Grid Forums
Category: Community
Forum: Hardware Chat Room
Thread: Raspberry Pi 4 8GB out! $75 (May 2020) |
| No member browsing this thread |
|
Thread Status: Active Total posts in this thread: 26
|
|
| Author |
|
|
hwierzbicki
Advanced Cruncher Joined: May 1, 2016 Post Count: 55 Status: Offline Project Badges: 
|
The Turing Pi is an interesting product if you're looking into compute clusters.
----------------------------------------Don't think the Compute 4 will end up being that much less than the standard Pi 4, market is much more limited and you gotta factor in the price of onboard EMMC. I'm mostly concerned about the heat dissipation for the Compute 4 modules, especially if they're crunching BOINC 24/7. Wonder if they'll be underclocked.  |
||
|
|
poppinfresh99
Cruncher Joined: Feb 29, 2020 Post Count: 49 Status: Offline Project Badges: 
|
The Compute Module doesn't have Ethernet or Wifi (or Bluetooth). It is designed to be used within a custom designed host board, which can supply all power and all connections via a SO-DIMM connector.
----------------------------------------https://www.raspberrypi.org/forums/viewtopic.php?t=124542 Since I don't have the skill to construct such a host board for a rack (I certainly couldn't do it cheaply), I was thinking about a Pi 4A (which would likely remove Ethernet, most of the USB, and some of the RAM of Model B), but there probably won't ever be a 4A... https://www.raspberrypi.org/forums/viewtopic.php?t=247927 So the best Pi for the foreseeable future seems to be the 4B! Edit: Oh, I just saw the Turing Pi post. Interesting. An advantage here is you only need to buy 1 power cord and 1 Ethernet cable. As for cooling, there would need to be a fan and something to guide air between the Pis. This is a disadvantage as this would take power. Oh, and a heat sink for each should probably be used? I really like the ODroid-N2! Powerful, energy efficient, has Ethernet, metal-housing heatsink, and microSD slot for storage. Has a DC power jack (12V and 2A power adapter is recommended). Is anyone running BOINC on one of these using Ubuntu (not Android)? I wonder how the big.Little architecture would work on Linux (background tasks on Android seem to use just the little cores). If BOINC just runs on the little cores, then I maybe wouldn't care for the ODroid-N2. [Edit 2 times, last edit by poppinfresh99 at Jun 9, 2020 6:09:52 PM] |
||
|
|
hwierzbicki
Advanced Cruncher Joined: May 1, 2016 Post Count: 55 Status: Offline Project Badges:
|
Oh, I just saw the Turing Pi post. Interesting. An advantage here is you only need to buy 1 power cord and 1 Ethernet cable. As for cooling, there would need to be a fan and something to guide air between the Pis. This is a disadvantage as this would take power. Oh, and a heat sink for each should probably be used? Not much of a disadvantage - I'd recommend an active HSF for any RP4 you plan to run BOINC on. You'll be hitting that thermal throttle pretty quickly if you don't. I really like the ODroid-N2! Powerful, energy efficient, has Ethernet, metal-housing heatsink, and microSD slot for storage. Has a DC power jack (12V and 2A power adapter is recommended). Is anyone running BOINC on one of these using Ubuntu (not Android)? I wonder how the big.Little architecture would work on Linux (background tasks on Android seem to use just the little cores). If BOINC just runs on the little cores, then I maybe wouldn't care for the ODroid-N2. Haven't tried Linux on my ODroid N2, but it's able to crunch on all 6 cores on Android. |
||
|
|
poppinfresh99
Cruncher Joined: Feb 29, 2020 Post Count: 49 Status: Offline Project Badges:
|
I'd recommend an active HSF for any RP4 you plan to run BOINC on. You'll be hitting that thermal throttle pretty quickly if you don't I Googled HSF---means Heat Sink and Fan. I have some numbers if anyone is interested. From my experience (never using a case and keeping Pi vertical) a little heat sink reduced temp by around 5 °C, but using a very low power fan reduced temp by another 14 °C. This 12V fan was being run from the Pi's 5V, so it was only using 0.17 watts (instead of 1 W), and wasn't snug with the Pi as they were just sitting right next to each other. I no longer have the fan (I borrowed it; my Pi 3B running 4 tasks now throttles but only rarely), but I'd be curious if the effects of fan and heat sink add or "multiply". Haven't tried Linux on my ODroid N2, but it's able to crunch on all 6 cores on Android. Very interesting, but being able to set "use 6 cores" in BOINC doesn't mean that all 6 are being used. Instead, it means that 6 processes are started. The following has been my experience on my big.Little Android phone... https://boinc.berkeley.edu/forum_thread.php?id=12532 Is your ODroid-N2 running Boinc as a background or foreground process? For my big.Little phone (4 big and 4 little), I can always set BOINC to "run 8 cores", but ADB shell (using top then ps commands) reveals that the usual BOINC mode of background only uses the 4 little cores, and, if I trick BOINC into running as a foreground process, it will use 3 little cores and 4 big cores (an 8th task would hop between little and big cores) https://www.xda-developers.com/quickly-install-adb/ In your case, if you have 6 cores where 2 are little, the WCG website would have the elapsed time be 3 times larger than CPU time if only the little cores are being used but BOINC is set to "use 6 cores". https://www.worldcommunitygrid.org/ms/viewBoincResults.do?filterStatus=4 [Edit 1 times, last edit by poppinfresh99 at Jun 10, 2020 5:22:19 AM] |
||
|
|
hwierzbicki
Advanced Cruncher Joined: May 1, 2016 Post Count: 55 Status: Offline Project Badges:
|
I Googled HSF---means Heat Sink and Fan. I have some numbers if anyone is interested. From my experience (never using a case and keeping Pi vertical) a little heat sink reduced temp by around 5 °C, but using a very low power fan reduced temp by another 14 °C. This 12V fan was being run from the Pi's 5V, so it was only using 0.17 watts (instead of 1 W), and wasn't snug with the Pi as they were just sitting right next to each other. I no longer have the fan (I borrowed it; my Pi 3B running 4 tasks now throttles but only rarely), but I'd be curious if the effects of fan and heat sink add or "multiply". The RP4 has a totally different SOC than the RP3 - you can't really compare results. I have an Armor Case for my RP4 and my temps are consistently ~40C above ambient. ETA Prime (YouTuber) is a good reference if you want to see how see how well active HSF and passive cooling cases work for the RP4 (hint: passive cooling isn't enough if you want to avoid the thermal throttle). Very interesting, but being able to set "use 6 cores" in BOINC doesn't mean that all 6 are being used. Instead, it means that 6 processes are started. The following has been my experience on my big.Little Android phone... https://boinc.berkeley.edu/forum_thread.php?id=12532 Is your ODroid-N2 running Boinc as a background or foreground process? For my big.Little phone (4 big and 4 little), I can always set BOINC to "run 8 cores", but ADB shell (using top command) reveals that the usual BOINC mode of background only uses the 4 little cores, and, if I trick BOINC into running as a foreground process, it will use 3 little cores and 4 big cores (an 8th task would hop between little and big cores) https://www.xda-developers.com/quickly-install-adb/ In your case, if you have 6 cores where 2 are little, the WCG website would have the elapsed time be 3 times larger than CPU time if only the little cores are being used but BOINC is set to "use 6 cores". https://www.worldcommunitygrid.org/ms/viewBoincResults.do?filterStatus=4 You can view which cores are active using apps such as CPU Monitor. Not sure why people are running into issues with big/LITTLE cores, I'm running BOINC on both my N2 and a OnePlus 6T with all cores active. Elapsed time for both is usually within 10-20 minutes of CPU time. |
||
|
|
poppinfresh99
Cruncher Joined: Feb 29, 2020 Post Count: 49 Status: Offline Project Badges:
|
The RP4 has a totally different SOC than the RP3 - you can't really compare results. I have an Armor Case for my RP4 and my temps are consistently ~40C above ambient. ETA Prime (YouTuber) is a good reference if you want to see how see how well active HSF and passive cooling cases work for the RP4 (hint: passive cooling isn't enough if you want to avoid the thermal throttle). I was trying to figure out for, in the Turing Pi setup, if it would be best to just put heat sinks on every Compute Module, just use a single fan for all of them, or use both sinks and fan. I suppose fans have initial cost + continued power cost. I would imagine that the laws of physics (thermal convection and conduction) are the same for different hardware. The relative effects of fans vs. heat sinks should then also be similar. Is this one of the video(s) you mentioned? https://www.youtube.com/watch?v=Wwk4HR9G51Y Seems that their results matched my results quite nicely. Like me, they also never test fan-only. Interestingly, fan being at 1 W did about the same cooling as mine being at 0.17 W. I suppose this makes sense because wind chill "has a larger increase" at lower wind speeds (that is, doubling wind speed does much less than double the wind chill)... https://en.wikipedia.org/wiki/Wind_chill#/media/File:Wind_chill.png You can view which cores are active using apps such as CPU Monitor. Not sure why people are running into issues with big/LITTLE cores, I'm running BOINC on both my N2 and a OnePlus 6T with all cores active. Elapsed time for both is usually within 10-20 minutes of CPU time. Oh awesome! Thanks for the link! I had tried many apps that never worked. I could have sworn I tried your app before (on my always non-rooted phone). It works! |
||
|
|
hchc
Veteran Cruncher USA Joined: Aug 15, 2006 Post Count: 865 Status: Offline Project Badges:
|
cwierzbicki said:
----------------------------------------The Turing Pi is an interesting product if you're looking into compute clusters. Don't think the Compute 4 will end up being that much less than the standard Pi 4, market is much more limited and you gotta factor in the price of onboard EMMC. I'm mostly concerned about the heat dissipation for the Compute 4 modules, especially if they're crunching BOINC 24/7. Wonder if they'll be underclocked. Ouch, the Turing Pi is pricey. $189 total / 7 slots = $27/slot. I don't think there is any cost savings going with the Compute Module 4 when it comes out over just buying the full RPi4 to be honest. C4 Labs makes an 8-slot "cloudlet" cluster case for $70: https://www.c4labs.com/product/8-slot-stackab...-computers-color-options/ I could get 7 RPi 4Bs (2 GB RAM or 4 GB RAM for more future proofing), an 8-port switch, and some small heat sinks to put on each of the Pis. The cluster case includes 4 active fans, so I wouldn't need to actively cool each individual Pi with its own fan. Another option I'm thinking about is I have an old Rosewill 2U rack-mount case I was using for my firewall/router (pfSense). It has room for three 80mm fans in the front, and I could make a bunch of slots to mount a ton of Pis vertically inside the chassis. And there would be room to hide an 8-port or 16-port switch inside along with all the necessary power cables. And simply route one AC cable and one ethernet cable out the back of the case -- for all anyone knows it's just a regular 2U server, not a cluster of a ton of SBCs. I think ultimately I need to put together an Excel spreadsheet and look into how much a cluster of Odroid N2s or a cluster of Raspberry Pi 4Bs would cost and find out the break-even point compared to an AMD CPU. I could throw in variables for cost, power consumption, performance/Watt, etc. At some point it may be cheaper to just build a single crunchbox instead of a cluster of SBCs. For example, the Intel Atom C3955 was released in 2017 and is 16 cores at 32W TDP. MSRP is pricey at $434 and server motherboards are very expensive. But surely at some point it may be cheaper to go with a power-efficient Intel or AMD crunchbox that is fairly cheap and has good performance/Watt.
|
||
|
|
poppinfresh99
Cruncher Joined: Feb 29, 2020 Post Count: 49 Status: Offline Project Badges:
|
C4 Labs makes an 8-slot "cloudlet" cluster case for $70: https://www.c4labs.com/product/8-slot-stackab...-computers-color-options/ I like how it takes care of air flow for you. I really like the cost! Instead of an Ethernet switch, would WiFi have any real disadvantages? Interference? Losing connection? Power usage? cwierzbicki might disagree, but one thing I liked about the Turing Pi is the simplicity of a lack of cables. For the C4-Labs thing, you'll also need a power adapter with 8 USB-C cables. Maybe something like the following 60W device and cables (for $50 total, and the cables are the most of it)? - https://www.amazon.com/Charging-Technology-Gu...d-Multiple/dp/B07XXDS86V/ - https://www.amazon.com/Braided-Compatible-Sam...le-Charger/dp/B07HQHL6ZK/ Edit: the first link is crap. I just noticed that the max current is reduced for the lower ports. Before I'd consider it, I want to know the average CPU time of cwierzbicki's N2 on an Open Pandemics task (no overclocking running max cores)... - Pi 3B: ~20 hours - Pi 4B: ~10 hours - ODRoid-N2: ??? hours <-- and also has 50% more cores! [Edit 1 times, last edit by poppinfresh99 at Jun 13, 2020 1:39:42 AM] |
||
|
|
KodeX
Advanced Cruncher Germany Joined: Aug 17, 2006 Post Count: 96 Status: Offline Project Badges: 
|
Before I'd consider it, I want to know the average CPU time of cwierzbicki's N2 on an Open Pandemics task (no overclocking running max cores)... - Pi 3B: ~20 hours - Pi 4B: ~10 hours - ODRoid-N2: ??? hours <-- and also has 50% more cores! I am running 5x Odroid N2 on Android atm. They take about 7.5-9.5h for an OPN WU (all 6 cores running 100% 24/7). The range of 2 hours might be because of different performance of big/LITTLE cores. Fastest WU and slowest WU of the last days. For the slowest WU the time is 9.59h compared to 35.75h of my wingman  [Edit 3 times, last edit by KodeX at Jun 10, 2020 5:08:09 PM] |
||
|
|
hwierzbicki
Advanced Cruncher Joined: May 1, 2016 Post Count: 55 Status: Offline Project Badges:
|
Ouch, the Turing Pi is pricey. $189 total / 7 slots = $27/slot. I don't think there is any cost savings going with the Compute Module 4 when it comes out over just buying the full RPi4 to be honest. Yep, going with a solution like the Turing Pi most likely won't save you money - it's mostly just for easy of use. Set it up, throw it in a cheap ITX SFF PC case with a couple of fans, and you're all set. No need to buy a switch or power supply/supplies (like poppinfresh mentioned). Price honestly isn't that bad, though, when you consider that the compute dev board the RPi foundation offers goes for $125... C4 Labs makes an 8-slot "cloudlet" cluster case for $70: https://www.c4labs.com/product/8-slot-stackab...-computers-color-options/ Been looking at that myself! Seems like an awesome kit. I think ultimately I need to put together an Excel spreadsheet and look into how much a cluster of Odroid N2s or a cluster of Raspberry Pi 4Bs would cost and find out the break-even point compared to an AMD CPU. I could throw in variables for cost, power consumption, performance/Watt, etc. At some point it may be cheaper to just build a single crunchbox instead of a cluster of SBCs. For example, the Intel Atom C3955 was released in 2017 and is 16 cores at 32W TDP. MSRP is pricey at $434 and server motherboards are very expensive. But surely at some point it may be cheaper to go with a power-efficient Intel or AMD crunchbox that is fairly cheap and has good performance/Watt. My guess is that ARM will win in the core/dollar and performance/watt categories. You can buy 6 2GB Raspberry Pi 4s (24 cores total) for a little over $200. Even after you throw in all the accessories, you're probably looking at a total build price of $350. And we already know that the RP4 has a better FLOP/Watt ratio than a 3900X (though that might change if you put the 3900X into 65W eco mode). That said, a AMD 3900X build would definitely win in the performance/dollar category. I know that Atom LOOKS good, but that's a lot of money to be spending on a device with a per-core performance not all that different than an ODroid N2. Instead of an Ethernet switch, would WiFi have any real disadvantages? Interference? Losing connection? Power usage? Probably depends on your router and how many devices are currently to your network more than anything else. Hardwired is certainly the preferred method, but both my N2 and RP4 are currently running on a relatively unloaded WiFi network with no issue. Power usage should be minimal. For the C4-Labs thing, you'll also need a power adapter with 8 USB-C cables. Maybe something like the following 60W device and cables (for $50 total, and the cables are the most of it)? Better option would be a dedicated power supply, like this. |
||
|
|
|