BI on Proxmox - which Ryzen CPU, and how easy to pass through GPU?

[itm]

I'm speccing up a replacement for my VM host - migrating from ESXi to Proxmox. My three VM's are:

• Blue Iris running on Windows Server 2022 (10 cameras, direct to disk, event recording only)
• Lubuntu running file server, Home Assistant, Logitech Media Server and web server
• Windows 10 running a PVR server.

I'm considering a number of Ryzen CPU's for the new host, but am not sure how much grunt I really need to support the above workload. This is an example of some that I am weighing up:

• Ryzen 5 8600G
• Ryzen 7 7700X
• Ryzen 9 5900X.

I have 2 questions:

1. Which of the above would be the most sensible choice for my workload? (or should I be looking at something else)
2. If I go for a model with an onboard GPU, how easy would it be to pass through the GPU to the Windows VM running Blue Iris?

 

[quinti]

when you passthrough the GPU, your host can only be used headless/remote, remember that.
Before last week my config was:

• Windows server 2022 VM on proxmox
• 2 NICs, one for the switch with the cams, one for the LAN
• Ryzen 5600g, 6 threads for the BI VM, 16gb
• 10 different cams, annke/dahua/reolink, all on 3Kor 4K, substreams enabled
• CPAI 2.8.0 on the same VM, yolov5.net, small model size

with this config, CPU usage varied between 15 to 100%, depending how much motion.

Now ive installed a m2 Coral TPU and using the coral 2.4.0 module of CPAI. CPU usage went down to 30% max, power consumption about 20W lower. But the accurancy of the detections is much lower then before because there are no custom models to use for Coral actually. and ive more then 10% failed inferences, with yolov5 on CPU ive had 0%. But still experimenting....

CAUTION!!!
Use virtio drivers for the NIC! First i used E1000 driver and had a frozen BI system every day

 

[itm]

when you passthrough the GPU, your host can only be used headless/remote, remember that.
Before last week my config was:

• Windows server 2022 VM on proxmox
• 2 NICs, one for the switch with the cams, one for the LAN
• Ryzen 5600g, 6 threads for the BI VM, 16gb
• 10 different cams, annke/dahua/reolink, all on 3Kor 4K, substreams enabled
• CPAI 2.8.0 on the same VM, yolov5.net, small model size

with this config, CPU usage varied between 15 to 100%, depending how much motion.

Now ive installed a m2 Coral TPU and using the coral 2.4.0 module of CPAI. CPU usage went down to 30% max, power consumption about 20W lower. But the accurancy of the detections is much lower then before because there are no custom models to use for Coral actually. and ive more then 10% failed inferences, with yolov5 on CPU ive had 0%. But still experimenting....

CAUTION!!!
Use virtio drivers for the NIC! First i used E1000 driver and had a frozen BI system every day

Many thanks for the feedback - it gives me a useful benchmark. I'm leaning towards a Ryzen 5700G, which would shave a couple of hundred off the cost, but should hopefully support my workload. I don't use AI for motion detection (which may be just as well as I've been reading many tales of failed attempts to passthrough Ryzen on-board GPU's).