The requirements depend a lot on the application. I understand you want a GPU to run Open Pandemics. Little is known about that application, the first observations are just trickling in. You should wait until you know more. In the meantime you could look if you have a CPU with integrated GPU available. Try to use that for a start and make your own experiences.


Nvidia's GT GPUs are display devices with some computing capability. I wouldn't buy any of those specifically for computing. Well, you could try a 1030 if you really want low power but understand that low power also means little output. You could soon find that you want more and have to buy twice.
My personal sweet spot used to be the low end GTX cards. I would consider a 1650 [Super] or 1050 [Ti], but not at the current prices. Another reason to wait.


Memory throughput can be a limiting factor. DDR3 is slower than GDDR5. Moreover, all those cards you mentioned use 64 bit memory, a few 710 even 32 bit. If the application is memory limited that could be a major drawback. But it could also be a little matter.
1GB is risky, you could find the card to be useless with that. Either immediately, or possibly after the next application upgrade or with larger tasks. I wouldn't buy less than 2GB and even that is not safe when you don't know what to expect.
I personally prefer passive cooling if I can get it. Less noise and no risk of mechanical failure. You just need some air flow in the case. All types you mentioned are available in passive cooled versions, just make sure the manufacturer hasn't crippled the card to allow for passive cooling, like underclocking or such. But I wouldn't expect that with cards that are low power anyway.
Also, your choices are mostly low profile cards. I'd specifically avoid LP cards with active cooling, those little fans can be noisy.

Oh, almost forgot your question about specs to look for. My approach is to first look at the GFLOPS rating. Single precision (FP32) is what usually matters but that also depends on the application. Then compare that against the TDP and price. You'll likely find that the small GPUs are not more efficient.
Rule of thumb: Less than 1000 GFLOPS is particularly slow, less than 500 GFLOPS is a toy.

Thanks Uplinger for pointing me to the log messages of the Boinc manager. I forgot to look there in the first place. I'm afraid my intel integrated gpu only supports OpenCL 1.1. As you can see I'm running an old Xeon. Being a server type processor I didn't expect it to have a great integrated gpu anyway.

Also thanks to all the others for your kind tips and explanations. It's very helpful.
SNURK

Starting BOINC client version 7.2.47 for windows_x86_64
log flags: file_xfer, sched_ops, task
Libraries: libcurl/7.25.0 OpenSSL/1.0.1 zlib/1.2.6
Data directory: C:\ProgramData\BOINC
Running under account Home
OpenCL: Intel GPU 0: Intel(R) HD Graphics (driver version 9.17.10.2932, device version OpenCL 1.1, 1496MB, 1496MB available, 17 GFLOPS peak)
OpenCL CPU: Intel(R) Xeon(R) CPU E3-1265L V2 @ 2.50GHz (OpenCL driver vendor: Intel(R) Corporation, driver version 1.1, device version OpenCL 1.1 (Build 30316.30328))
Processor: 8 GenuineIntel      Intel(R) Xeon(R) CPU E3-1265L V2 @ 2.50GHz [Family 6 Model 58 Stepping 9]
Processor features: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss htt tm pni ssse3 cx16 sse4_1 sse4_2 popcnt aes nx lm vmx smx tm2 pbe
OS: Microsoft Windows 7: Ultimate x64 Edition, Service Pack 1, (06.01.7601.00)
Memory: 5.70 GB physical, 11.40 GB virtual
Disk: 111.69 GB total, 53.56 GB free

Hi sam6861,
I thank you for the provided figures.
On my side, since I am not a gamer, I have very limited knowledge regarding GPU and I appreciate some inputs.
Regarding the performance you reported (GT 1030 vs RX 550), is it on Windows or Linux machines?
Cheers,
Yves

Here is a good place to start: BOINC's own GPU Computing. In your quest you will encounter some facts, some myths, and diverse opinions. I don't know enough about the WCG GPU client to know which GPUs will be best. Does it use FP64 or FP32? I have not encountered a GPU client before that measured time in tenths of an hour instead of seconds.

Here is another article: BOINC Client configuration.

Use a search engine to learn about app_config.xml. The app during the current beta is beta29. You will have to go to File Folder Options | View tab | select "Show hidden files, folders, and drives" in order to see Program Data.

I use AMD GPUs, so I am familiar with the use of Radeon Control Panel and MSI Afterburner. They are useful tools for monitoring GPU useage, tuning power, fan speed and VRAM speed. I have not encountered a GPU client that required high video RAM speeds, so I run it as slow as possible to reduce heat. On my AMD Radeon HD 7990s, I run +10 Power, 150MHz VRAM speed, and up to 75% fan speed to control temperatures. For your GPUs you will have to find settings that work best for you. I do not recommend overclocking because the loads already produce high temperatures.

For a short-running client, which this one appears to be, users will run out of work units in queue and wait before more work units are downloaded. If that is the case you can find boinc.exe version 7.15.0. (Windows) as developed by JStateson for short-running work at Milkyway@Home. There's a .deb user's version, too. Follow the Read Me. You are simply replacing your current boinc.exe in Program Files/BOINC (Windows). Your installation of Windows needs to have Microsoft Visual C++ in your list of Apps. You must exit the BONC client before installing the new boinc.exe. After a restart, your BOINC Manager will be version 7.15.0.

Happy crunching!

@rickjb
Nice!
If you run opng, you will need to run multiple units at a time to load that card. 1 unit only had my rtx 2060m at 26%. 4 at a time was 50 something %, so it does not scale up linearly adding the work units. I did not try more than that (or gather good data really), since they only allowed 4 to be assigned at a time on the last beta and it was a short run.

Personally, I would like it if it was designed to utilize the card more (without me having to add work units until it looks good). For comparison, folding at home averages 92-98% depending on the project. To be fair though, they can and do have different graphics card requirements per project and assign work based upon work unit availability and what gpu is in the client. I think that is probably beyond the complication level that WCG is setup for.