O3 almost not using any GPU power on a Mac
![[AF>Le_Pommier] Jerome_C2005](https://einsteinathome.org/sites/default/files/imagefield_default_images/default-avatar.png?1347855344 "[AF>Le_Pommier] Jerome_C2005"){kind=avatar}
23 Apr 2022 10:01:42 UTC
Topic 227438
(moderation:
Hi
I have one of the lastest iMac Intel (end 2020), on it the application "Gravitational Wave search O3 All-Sky #1 v1.01 (GW-opencl-ati)" is hardly using the GPU (only a few % of GPU power). I have noticed this long ago and removed the app from my setting, recently I gave it another try but same thing : after one hour of computation I can see GPU history used below 5%.
So I crunch on "Gamma-ray pulsar binary search #1 on GPUs v1.17 (FGRPopencl-ati-mav)" (this is the only application that Einstein is sending to my mac, though the others are activated in my parameters, except O3 again), which is using my AMD Radeon Pro 5700 XT GPU pretty well, but I can pile up 3 tasks to maximise utilisation (I get 80% average GPU usage with 3).
But sill I wonder : am I the only one to see such blatant ineffectiveness with the GPU for the O3 application ? is it linked to this GPU / family of GPU ? or macOS ? or the combo of both ?
Thanks for your comments.
Hi Fellow Mac Cruncher!I
)
Hi Fellow Mac Cruncher!
I crunch FRGPopencl-ati-mav tasks on i5 hex-core mac minis (Mac Mini 8,1) connected via Thunderbolt 3 to eGPUs containing AMD RX 580s, RX Vega 56s, RX Vega 64s, RX 5700 XTs, and RX 6600 XTs. None of those GPUs are as productive as comparable GPUs in Linux or Windows boxes. I suspected it was due to the limited bandwidth of the Thunderbolt 3 connection. But if your 5700 XT is an internal GPU, your apparent lack of productivity makes me think it is an Apple OS issue.
Regards, David
"I was born in a small town, and I live in a small town." - John Mellencamp
Yes this is an iMac with
)
Yes this is an iMac with internal stuff (CPU, GPU).
I only have external a 2 GB SSD (given Apple price...) with TB and it is quite effective (I chose the smallest internal SSD for my config). I also have chained old storage devices (TB3 --> TB2 --> FW800 if I remember well, I had to buy 2 adapters ) but this is mainly for backup and speed is not a concern.
But you seem to have a hell of an external box, with 5 GPUs !! have you tried to use only one at a time / deactivate the others to see if it makes any difference ?
Also when you say "None of those GPUs are as productive as comparable GPUs in Linux or Windows boxes" what do you mean : below 5% of GPU usage like me, of "a fair amount of use but not maximized" (e.g. 50%, or even 30 or 20%) ?
Sorry. I should have been
)
Sorry. I should have been more explicit. My E@H crunchers include a MacMini 8,1 connected to 2 RX 580s, an RX Vega 56, and an RX 5700 XT; a MacMini 8,1 connected to an RX 580, an RX 6600, and an RX Vega 64; and a MacMini 8,1 connected to an RX 580 and an RX 6600 XT. E@H is the only BOINC GPU app that I run on these computers and these computers run BOINC ~90% of the time. I can estimate the observed E@H RAC (under best conditions) for each GPU on each of these computers and these RAC values are much less than half of the observed E@H RAC values of identical GPUs running in (non-Mac) Windows or Linux boxes.
I have three other MacMini 8,1 computers connected to a total of 10 AMD GPUs. Collatz at Home is the only BOINC GPU app that I run on these computers and these computers run BOINC ~90% of the time. I can estimate the observed C@H RAC (under best conditions) for each GPU on each of these computers and these RAC values are comparable to the observed C@H RAC values of identical GPUs running in (non-Mac) Windows or Linux boxes.
So, based on both of our experiences, it appears that the E@H software does not play well "with" MacOS or Mac hardware, regardless of whether the GPU is internal or external. By the way, my eGPUs get much hotter when crunching C@H tasks than when crunching E@H tasks. This is consistent with the hypothesis that C@H does a better job in maximizing the utilization of eGPUs than does E@H.
"I was born in a small town, and I live in a small town." - John Mellencamp
Your case is very particular,
)
Your case is very particular, having so many external GPU connected to one machine (up to 4 different GPU for the 1st macmini if my count is correct), great for you that is seems to work, I didn't even know boinc could handle configuration like this, and work well.
You seem to only rely on a high level credit output comparison to have an idea of the production effectiveness of these configurations, I'm not sure if you can really know how does each one of these GPU behave for each type of application they get some tasks to crunch... but now when you have so many of them, it's probably not that important :)
I was really talking about looking in real time for the GPU usage (% of max power used), for me making sure it is actually used is important, since I have only one.
And clearly on my mac, the opencl-ati O3 application is almost not using any GPU power.
[AF>Le_Pommier wrote:
)
Each application is written to get the best possible result and sometimes that means using ALOT of cpu power as opposed to gpu power when it has too, meaning sometimes the gpu's don't get much usage even though the tasks do run much faster on the gpu then a straight cpu app.
Thanks for your answer :
)
Thanks for your answer : do you mean that "for a given application (in this case O3) GPU usage will vary from one task to another one" ? or your answer is more globally for one application ? can it vary depending on the OS of the application ? (in my case macOS)
Because my experience was (hardly) "no GPU usage" (monitoring in detail) during all execution time of the O3 tasks...
So, as I explained above, I only use GRP which is actually using GPU on my machine - I prefer to let GPU task for projects / applications that actually use the GPU, I feel like it is a waste of resources if I don't :)
I don't know what percentage
)
I don't know what percentage of my GPU's capacity is utilized by the S@H GRP tasks. I suspect it is less than utilized by Collatz@Home GPU tasks under optimized parameters, as those tasks make my GPUs run much hotter than the S@H GRP tasks. But, there are not many BOINC projects available for Macs with AMD GPUs, unless you are willing to run Linux or Windows on those boxes. And, it's only getting worse, as I am not aware of any BOINC projects that support the Apple Silicon GPUs.
By the way, I believe your iMac has two Thunderbolt 3 ports. If so, it can support multiple eGPUs. If you use Akitio Node Pro eGPU enclosures (the model with a pass-through TB3 port), you could theoretically connect up to 4 eGPUs to your iMac. I have not done so with your model, but I have done so with a dual-core i5 20" iMac. However, you have to be careful about the thermal load. Those iMacs don't disperse heat very well.
"I was born in a small town, and I live in a small town." - John Mellencamp
This is NOT Seti@home
)
This is NOT Seti@home
Tigers_Dave wrote: I don't
)
One is enough for me :D (and I don't even think about the cost about such a setup)
You can know the % of GPU used : in the standard task monitor of macOS on each detailed process you get both CPU and GPU usage (at least on recent versions of macOS).
Then tools like iStat Menus (highly recommended) give you a global overview of the GPU usage (this is how I know that I have to pile up to 3 GRP to maximize the usage).
Keith Myers wrote: This is
)
??