What card can help with deepstack?

S

Schrodinger's Cat

Young grasshopper
Nov 17, 2020
54
23
USA
I'm currently running 4 cams, Intel 4th Gen in a SFF desktop box (no full sized cards). Windows CPU deepstack pegs my CPU to 100 and take 500-1000 ms for analysis. The Docker version is much better (sub 400ms and no CPU pegging), but I can't integrate that directly with Blue Iris so I'm currently using AITool.

Say I want to ditch AITool and move over the Windows GPU version for direct BI integration. How much do I need to spend on a card for good performance? Thanks
 
There are a few threads discussing this exact scenario. I believe a P400 card would fit in the SFF and would work well enough. The SFF kind of limits things since you can't get a full height, higher horsepower, GPU into it. You'll still see CPU spikes as python is detecting, but the should be reduced. A 4th gen CPU is kind of long in the tooth of todays world and for DS.
 
  • Like
Reactions: looney2ns
Schrodinger's Cat said:
Which 1030?
Click to expand...

sub 100ms?? woww.

Just with that 110 USD card ? (unfortunately no DP outlets on that though - I have a cheap amd firepro with 4 x minidp for my 3 screens for work but use the intel cpu for deepstack). Wish deepstack could use amd gpu.
How much difference would a 4gb geoforce make over the 2gb one?

What cofig settings do you have for the cams? (fps/resolution/size etc).
Also is that on deepstack MODE High/Medium?

I have an intel i7 7700 with 32gb ram. 300ms is about the average for a 768x432 in mode High.
 
Last edited:
Update: I bought the card linked. On medium settings I'm getting around 150ms, on high about 250ms. This is much better than before, I am happy. My old 4th gen now sits happily between 10% and 20% CPU almost all the time, might go to 30% on detections. I expect to get at least a couple more years out of the old girl. Thank you.
 
  • Like
Reactions: Pentagano and sebastiantombs
Awesome! I have been playing with the mode, and at least in my field of view I have been able to change mode to LOW and still get high % confidence at sub 100ms. I have also moved some of the camera HA over to the GPU as well.

It isn't a screamer GPU, but it works in the 4th gen and takes just enough off to get more availability out of the machine!
 
  • Like
Reactions: Flintstone61 and sebastiantombs
As an option for anyone else, I have my main desktop computer running Deepstack server and the Blue Iris PC queries it over the local network. Most detections are between 100-150 ms. I do have some under sub 100. RTX 2070 on my desktop. I am probably going to run it this way until the prices drop.
Mode=High
 
  • Like
Reactions: Markles, sebastiantombs and Flintstone61
The RTX-2070 is a far more capable card than a GT-1030. About six times the number of cuda cores and triple the memory.
 
What would be the most affordable card to achieve ~100-200ms with mode high? Low profile for my Dell 3420.

Would a gtx 1030 2GB make any improvement over a i7 7700 cpu? Get around 300-350ms per image currently.
Or does more money need to be spent on at least a 1050?
 
Last edited:
Pentagano said:
What would be the most affordable card to achieve ~100-200ms with mode high? Low profile for my Dell 3420.

Would a gtx 1030 2GB make any improvement over a i7 7700 cpu? Get around 300-350ms per image currently.
Or does more money need to be spent on at least a 1050?
Click to expand...

I'm getting almost that with the $110 1030 linked above. I'm sure you can spend a bit more and do better, but this is good enough for me. Certainly way better than on CPU, but I'm running an i5 4590.

For awhile I was running the Docker version of deepstack and getting about 350 on medium. Limited to 2 cores. Have you tried that? It performed much better than the Windows CPU version which was giving me 1,000 ms and 100% CPU spikes.
 
  • Like
Reactions: sebastiantombs
Oddly enough I've just shut down my docker desktop versions. Yes had them limited to a couple of cores also. Used them for about 10 months.
Found the vmmem was taking up too much cpu so was offset by that.
Turned off docker etc and use windows version now
Works well.
Idles at 5 or 6% now. 10% less.
Doesn't max out either.

All depends what Gen and threads you have I guess. Also reinstalled the windows cleaned out all the bloat which always helps
 
Biggest improvement I made was illuminating my gardens with led lamps and keeping the cams in colour. Before I'd have the cpu working overdrive all night when I we had fog or heavy rain with the IR.
Now almost no false triggers.
 
Haha, working overdrive is basically my camservers life. I live on a busy downtown street in a top 25 city in the US with constant pedestrians all hours of the day who's shadows cast across my driveway, with my entire property under a canopy of deciduous trees in a place that gets lots of wind and almost 300 days a year of sun. Then at night it's constant headlights sweeping across my entire front yard.

My detection mechanism is basically going brrr 24 hours a day. I average over 5,000 images sent to deepstack every day! If I turn down my motion settings I start missing things. This AI thing was a game changer.
 
Schrodinger's Cat said:
Haha, working overdrive is basically my camservers life. I live on a busy downtown street in a top 25 city in the US with constant pedestrians all hours of the day who's shadows cast across my driveway, with my entire property under a canopy of deciduous trees in a place that gets lots of wind and almost 300 days a year of sun. Then at night it's constant headlights sweeping across my entire front yard.

My detection mechanism is basically going brrr 24 hours a day. I average over 5,000 images sent to deepstack every day! If I turn down my motion settings I start missing things. This AI thing was a game changer.
Click to expand...

Omg. Hope you have decent cooling on your cpu.
Luckily we're on a quiet street also with a high fence at the front. Only if someone uses the outside part of the driveway to do a u-turn then the light change will trigger a few snapshots.
 
You must log in or register to reply here.
Top Bottom