Blue Iris

Hello,

I belive I need to add a GPU to my system. AI is taking about 2-3 secs to detect objects. I do direct to disc recording and everything else is tweaked on the machine. Any suggestions on a cheap GPU that would work?

I have a Vostro 3671
24 GB RAm
Intel 9th Generation Core i5-9400 (9 MB Cache, 6 cores, 6 thread, 65 W, 2.9 GHZ
1 Seagte HD for Operating system
6TB WD Purple for deicated camera storage


Thanks I really appreciate it,
Michael

What you get will be dictated by your budget. You can run fine with something like a 1080i-level or equivalent card, or spend more and get a more current card, like a 3080-level performance.

Other people have had success with other cards and brands, maybe they can weigh in with some suggestions.

Thixotropic wrote: ↑Fri Sep 22, 2023 4:16 am What you get will be dictated by your budget. You can run fine with something like a 1080i-level or equivalent card, or spend more and get a more current card, like a 3080-level performance.

Other people have had success with other cards and brands, maybe they can weigh in with some suggestions.

Thank you @Thixotropic

Is the OS drive an SSD?

I’d have thought you’d be fine with that system. Way back when, when I tried Deepstack in a docker container I was getting response times of between .75 and 1.25 seconds in my rig which has a measly i5-6500 CPU.

How many cams do you have and what res are they running ?

My $0.02/:
Client's system. 13 cameras, similar to yours, don't remember model number.
1. The card must be supported by AI server and the system power supply.
2. Memory on the card is most important need, IMHO.
3. How big are the models you are using for search. How many item types in the search.
4. How many of your cameras are using AI and what is the probability of triggers close together.

I am running a lowly K620 (2GB) via docker on a minimal Linux VM that is in an old repurposed ESXi system. I don't do that many cameras with the AI, 4 at the moment. My need is mostly to overcome headlights at night, I don't look for anything other than persons and vehicles (General model). Typical reco times are under 100 msec, usually in the 60's, nothing found in the 50's. Challenge is vehicles hit one camera after the other. One camera is inside and watching a door for persons, false alerts due to insects (area is inside but not air conditioned and nothing to keep insects out).
I am looking to upgrade to a 1080 with 8GB to allow more cameras, as soon as the client approves. Alas, I don't think the power supply in the BI system will support it so will have to stick with the VM. I really don't like multiple failure points.

I have to allow for how many triggers at same time or close together.

I have been running AI in docker on VM's (3 similar systems) so as to easily upgrade/downgrade and change stuff without messing with the BI system. Fortunately I have 2 systems that run ESXi and have license for unlimited installations. ESXi allows pass through of the GPU, many of the VM servers don't. I run one BI system here, 3 cameras, mostly for testing, does watch my patio and front entrance. That system also has a K620. I purchased a Dell Precision from eBay that had 2 K620's in it, use it for ESXi with both the BI and Linux system. Took the other GPU to client's system. ESXi system makes it easy to have the camera LAN separate from business LAN and administer both from one system. I fully agree with Tim-the-Guu about keeping the camera LAN separate, many of the cameras want to phone home.

If your system is a production system, move slowly, test with clones.
You should probably look around on the Codeproject site if you haven't done so.


/end $0.02

Some good info there

As said above, do you use AI on all cameras, or only those that need help in removing false triggers ? There may be some tweaking to do.

If you have enabled Face or ALPR modules, then you may have inadvertently now got them working on ALL cameras which will seriously slug your cpu. The way to find out is to view the DAT files as you will see the modules being used per camera. I ended up turning these off on other cameras by using :0 like this in "Trigger/Artificial Intelligence":

Custom models: faces:0,alpr:0,ipcam-combined

The :0 forces the module OFF for that camera.

Then re-check DAT files to see they are no longer in use. Note: Turn DAT files OFF when you are done, else they will flood your Alerts folder

I personally use a GTX1650 which was suggested by MikeLud for use with his CPAI ALPR module. I also run the YOLOv5.NET module, as I found it to be faster than the Yolov5 6.2 model. YMMV.

AI: Default, or the faster custom models ?

Good evening everyone,

I want to thank everyone for their advice on this. To answer a few questions.

I have 10 cameras running at 15fps
AI set up on 3, Driveway Front Door and Backyard.
All I was trying to setup up AI for was to eliminate false alarms, mostly shadows and trees blowing, and show people, vehicles and animals.
With the advice you gave to check the .dat file I realized I had it setup for facial recog, lic plates, etc, what the hell, I didn't check those boxes in AI. What I did have checked was in the system setup to use Custom Models, I didn't realize it added everything.

I seem to be on the right track now just need a bit more tweaking. Hopefully you won;t mind if I come back for more advice. Have a greatweekend.

Thanks again!

Michael

Excellent, and you can ask as many questions as you like

Do let us know if this is enough to reduce your cpu load without a gpu.

TimG wrote: ↑Sat Sep 23, 2023 5:53 am Excellent, and you can ask as many questions as you like

Do let us know if this is enough to reduce your cpu load without a gpu.

Tim,

Thanks,

I ended adding a GPU late yesterday. For some reason 2 cameras would not use HA period on this computer. They did use HA on my test computer. Once I added a GPU late yesterday I could select HA on both cameras for CPU w/out issue.

Yes it has they both are staying consistant in single digits to low teens.

Michael