I have been testing VMIX on G4 instances on AWS also. I haven't done a live show yet but did notice audio drop outs every now and then
I just found this article on GPU monitoring so am going to test it out.
Thought I would share just in case it helped anyone else

https://aws.amazon.com/b...-with-amazon-cloudwatch/

I experienced a similar issue to this today, NDI audio would drop out for 1-2 seconds, maybe once a minute or so.

Found this thread while trying to diagnose it, and having solved it in our configuration thought I would share.

TLDR
Had a 2nd network adaptor running that we normally wouldn't use, for a separate LAN that was a one off for this production. Disabled the device in Windows and problem solved.

Details
vMix machine, ample resources in all respects, nowhere near any hardware bottlenecks.
8 vmix calls connected.
NDI outs 1 - 4 all in use, each a unique mix, going to various monitors and other destinations across network.
NDI outs for all 8 vmix calls in use, primarily to provide sound engineer with individual audio channels for each caller, but also various other destinations.
Various NDI ins, SDI ins, graphics, VTs etc.
All vMix audio busses in use

The core of this setup is a configuration we have been using extensively for months, and not experienced any NDI audio dropouts.

On this occasion, for various reasons we had decided to make use of the 2nd network adaptor in the vmix machine which is usually left disabled, for a 2nd separate LAN of a few devices we wanted to isolate from the main LAN.

This 2nd LAN only had a handful of devices with manually set IPs. No DHCP server, no gateway etc.

The 8 callers NDIs were being pulled over this LAN for audio.
Audio would drop out for 1-2 seconds, maybe once a minute or so.

Audio was unaffected at source in vmix.
Audio would dropout for every destination that was viewing the caller NDI feeds, on either of the LANs, and even on studio monitor on the local machine.

The 2nd LAN being the only difference with this setup to usual, that was the first place I looked. During a VT we quickly moved the sound engineers NDI receive back onto the main LAN as usual. No improvements.

I then completely abandoned the 2nd LAN, literally unplugged the ethernet cable. Still no improvements.

Now utterly confused, and pressure mounting from the client with the audience starting to comment on the audio dropouts, I took a major risk and disabled the ethernet adaptor in Windows, which is the state the device would normally be in.

Problem solved!

So there is probably a bug with NDI and multiple ethernet adaptors. Perhaps this is not a supported configuration for NDI anyway, yet to look into that.