There has been some discussion about threadripper cpu’s not being suitable for use with vmix.

However, if you want to do high channel count SDI, you really have no choice as you need the 64 PCIe lanes that threadripper boasts.

We decided to bite the bullet and go all out to make it work.

Results below, I’ll add screen shots shortly.

X399 Aorus Motherboard
Threadripper 1950x 16 core CPU with 32 logical processors and 64 PCIE lanes.
32Gb DDR 4
Gigabyte Aorus Nvidia GTX 1080TI Extreme Edition with 11Gb
512Gb Samsung NVME
2 x Decklink Quad 2 cards for a total of 16 SDI inputs
Intel SFP+ card with ubiquitous single mode fibre modules for 10Gbit networking


Results

In vmix at 1080p50 with all 16 SDI inputs running
5% Vmix CPU, 10% total, 10ms latency
GPU at 25%

We still have some more testing to do on end to end latency from SDI in to NDI out which I will post.

It’s mounted in a 4u rack and the plan is to use this as a standalone system for smaller jobs.
For larger jobs this will be the capture machine and will export all channels to a second machine over NDI which will do the mixing.
Because of the high channel count it needs 10Gbit so we went with fibre to allow us some flexibility on distance.

For example, at a concert we can have this machine on stage with all the cameras connected to it then run a single fibre out to our OB van where we do the mix. This keeps the setup nice and simple.

Let me know if you have any questions or tests you want me to run. We are pretty happy with the performance of this machine. It eats out intel 7700 machines for breakfast on high workloads like this and the high number of PCIe lanes means we could in theory add another decklink card for 24 inputs.

Hi Richard

We were running a countdown timer so you can measure latency.
Screenshots attached.

We are still doing a bit more testing with the decklink cards in different slots as there are a few dropped frames there. Normally we run in 25p so at this point we are really just stressing it to see if it breaks. It looks like the decklink cards are saturating their own 8 pcie lanes. I'll do some testing on this front to confirm.

some more digging and one of the decklink cards is sharing pcie lanes with an SSD so this is dropping frames and increasing latency on one of the cards.
We also tried changing to a cinema test pattern and it had no appreciable difference on load.
We tried pulling 16 NDI sources from it and it almost maxes out the GPU but it doesn't drop any NDI frames.
Again this is at 50p so a higher frame rate than we will actually run.

Interesting. What GPU? It occurs to me that the right choice of GPU would deliver unlocked nvenc that could push that further.

12 multicorders at 1080 50p is a lot!,

cpu at 100%?, what latency?.

What compresor and bitrate?. H264 at 75mbps?.

According to this: https://developer.nvidia...ecode-gpu-support-matrix

The GTX 1080 Ti has only two physical encoders, each restricted to one session.

Using a Quadro P2000 provides 2 physical instances of nvenc, without restriction of the number of sessions.

Using a Quadro P5000 provides 3 physical instances of nvenc, without restriction of the number of sessions.

Thanks a lot for sharing your system!

This makes me feel a lot more confident with my own leap of faith, although Im going for the 2990WX myself.

Indeed on the Multicorders maxing out GPU, the hardware encoders only really offload it. I wonder if you can get more going with a second video card.
I believe there are some older Quadros you can get with loads of Encoders as well for low price, if you were to build a pure recording device.

I went to IBC today and talked to Tim a bit about AMD, and the main reason for not testing on AMD was the PCI lane latency the first series of TR and Ryzen had. Your results reinforce my idea that they had it fixed in series 2, as this latency was also a major cause of bad gaming benchmarks for TR1.

Ive got the parts in today, so later in the next week ill be testing my system as well.

How would the lower spec Quadros perform for purely multicorder purposes? And can we use multiple cards to up the count?

How are these systems perfoming? Has anyone run their system with real world multiple NDI inputs into vMix?