RTX 4070 build advice for ~30× 8 MP Cameras on Proxmox ?

[mcooper1541]

RTX 4070 build advice for ≈30 × 8 MP Reolink on Proxmox 8.4.1

Seeking best-practice host/guest model, docker-compose flags, 0.16 timeline, Coral co-existence & snapshot strategy

1. Host / guest model

• Stay on privileged LXC or switch Frigate to a full Debian VM for cleaner NVIDIA isolation and easier snapshots/backups?
• Any gotchas with nvidia-container-runtime in LXC on PVE 8.4?

2. docker-compose flags & FFmpeg presets

Looking for configs that maximise NVDEC/NVENC on the 4070 but fall back to Intel QSV if the GPU is saturated.

• What ffmpeg: flags work best?
• Which env vars (NVIDIA_VISIBLE_DEVICES, NVIDIA_DRIVER_CAPABILITIES, etc.) actually matter?

3. v0.16 timeline

0.16.0-beta3 (tagged 2025-05-31) is out, but docs still list 0.15 as stable.

• If v16 is imminent I’ll wait; otherwise I’ll pin to 0.15.1.

4. RTX 40-series quirks

• Minimum driver version that’s known-good?
• Encoder/decoder session limits when record + detect + restream together?
• VRAM headroom for 30 × 4 K @ 15 fps?

5. Dual Edge TPUs — worth keeping?

I can still mount two Coral Dual Edge TPUs. Any practical gain off-loading detection to the TPUs while the 4070 handles decode/encode?

---

Current host summary

Component Details	Current Config
Server	Minisforum MS-01 (i9-13900H, 96 GB RAM)
Hypervisor	Proxmox 8.4.1 · kernel 6.8.12-11-pve
GPU 0	Intel Iris Xe (iGPU)
GPU 1	MSI RTX 4070
Accelerators	2 × Coral Dual Edge TPU (m.2 2280 B+M key)
Cameras	30 × Reolink 8 MP H.264 @ 15 fps
Frigate	Privileged Debian 12 LXC · Docker 25

Thanks in advance for any pointers!

[NickM-27]

1. Proxmox is not officially supported, Debian VM would align more closely to the recommended installation but if things are working well currently no reason to change
2. preset-nvidia
3. There is no planned date. There will be at least one more beta and then RC
4. Nvidia has a limit of 3 encoding sessions unless you do a driver patch. But frigate doesn't do any decoding by default
5. It will be simpler but the models that it can run are smaller. Depends on your priorities

[kirsch33]

For reference to the OP here's a link to the driver patch:

https://github.com/keylase/nvidia-patch