Yolo v5Net vs Yolo V5 6.2 on I7 without seperate GPU

[scoob8000]

From my understanding from the past, it was faster to run the .net version of yolo if you don't have a supported GPU.

My machine is a i7-6700.

With every update of code project, it seems to default to both running and possibly using both. I counter for .net never goes up, but I see log entries for both.

I just want to make sure I'm setup for optimal performance...

 

[MikeLud1]

From my understanding from the past, it was faster to run the .net version of yolo if you don't have a supported GPU.

My machine is a i7-6700.

With every update of code project, it seems to default to both running and possibly using both. I counter for .net never goes up, but I see log entries for both.

I just want to make sure I'm setup for optimal performance...

View attachment 178879

View attachment 178878

You should only have one Object Detection running so disable Object Detection (YOLOv5 6.2).

 

[scoob8000]

Gotcha!

 

[kaltertod]

In my experience with my 2 servers .net is faster. However, 6.2 is much more accurate in its identification.

 

[scoob8000]

In my experience with my 2 servers .net is faster. However, 6.2 is much more accurate in its identification.

At a rough estimate, how much faster?

A long time ago I tried taking a average to see which was faster, but back then .net was marginally faster.

I haven't done my test lately but at a glance my times are all over the place. 200ms to 800ms or so. Occasionally up over a second. Probably depending on camera and detail.

I was recently using ipcam-combined and delivery but ne

 

[kaltertod]

.NET server is using a geforce 1650

Average long term on my .net server using gpu 18-20ms

6.2 Server is using a 2060 TI 12GB

Average long term on that machine is running an average with gpu of 25ms with some spikes up into the 200ms range.

Running the 6.2 server in cpu mode nets an average of 250ms, with spikes up to 1000ms using and I7 5820k with a 4.4ghz core clock, That system also has 64gb of ram running 3600Mt/s

I have however not tested .NET in cpu mode.

Are you analyzing the sub stream with the AI or the mainstream? This will have a large effect on detection times.

**EDIT**CPU on .NET machine 125ms average with a Ryzen 5600X stock clocks, PBO enabled, 32GB of ram (This is kind of an apples to oranges comparison, might download .NET on my main server and run some tests with cpu only)

 

[scoob8000]

.NET server is using a geforce 1650

Average long term on my .net server using gpu 18-20ms

6.2 Server is using a 2060 TI 12GB

Average long term on that machine is running an average with gpu of 25ms with some spikes up into the 200ms range.

Running the 6.2 server in cpu mode nets an average of 250ms, with spikes up to 1000ms using and I7 5820k with a 4.4ghz core clock, That system also has 64gb of ram running 3600Mt/s

I have however not tested .NET in cpu mode.

Are you analyzing the sub stream with the AI or the mainstream? This will have a large effect on detection times.

**EDIT**CPU on .NET machine 125ms average with a Ryzen 5600X stock clocks, PBO enabled, 32GB of ram (This is kind of an apples to oranges comparison, might download .NET on my main server and run some tests with cpu only)

I'm just analyzing the substream. But holy cow is your performance better than mine, even in cpu mode.

I need to do some more testing. What model size are you using? small, medium, large, etc

 

[kaltertod]

I'm just analyzing the substream. But holy cow is your performance better than mine, even in cpu mode.

I need to do some more testing. What model size are you using? small, medium, large, etc

Large Model size

Background tasks on your server will have a large impact on performance. My machines are dedicated directly to BI, They run a few other apps like vnc server. The 5820k rig is running a dual monitor setup, the 5600x setup is running headless with no console running so that will contribute back some unused cpu cycles.

Are you hosting CPAI local on the BI Server?

try tuning your ram timings, and another thing that I use that seems to help is ISLC. I have it tuned just like you would with a gaming pc it seems to shave about 25-50ms off the analyze time. Both of my rigs are using 8tb spinners for storage but the DB's are on nvme ssds. The 5820k is also running quad channel on the ram config.