Red raw playback on a SuperMicro

I have a new SuperMicro system, and I need a little help troubleshooting some choppy playback, or wondering if this is even possible
The system is a 7048GR-TR with 2 18 core processors. This is hooked to a Cubix expander and I have 3 1080Ti GTX cards(will soon be 4), 1 in the PC and 2 in the Cubix. The hard drives (8 x 8TB) are connected via a 6Gb SAS connection and are in a raid 5, a BlackMagic Disk speed test yields around 1000 MB/s in read and write.
The plan is to be able to play a timeline back with 6k raw media on a UHD timeline. If I could play 1 track of video I would be happy, no expectations of multiple raw tracks.
Currently the playback is choppy. I set up a performance monitor for the GPU's and they only go up to 20% usage, cpu shows 80%, disk shows only 20% usage at 200MB/s max.
Is this most likely a response issue of the hard drives, or what is likely the playback bottleneck?

Thanks,

C

Decoding that footage is a CPU workload, and a very demanding one at that.

I did another test and optimized one of the raw clips to DNxHR HQX and dropped that by itself into a new timeline, optimization enabled, and it still had the choppy playback. I then tried moving the clip to an internal ssd drive and have the same issue.

[Dwaine Maggart]

Are you using Resolve Studio 14.2?

Do you have Performance Mode enabled in the Playback menu?

Does your Video Monitoring resolution match your timeline resolution? Try that, if you can.

If you can't, in Project Settings - Image Scaling, uncheck the Output Scaling "Match timeline settings" checkbox, and set the Output Resolution to match your Video Monitoring resolution.

Yes 14.2.

Performance Mode is enabled.

Yes, timeline is 3840 x 2160 and monitor is 2160 23.976(same as timeline).

I tried unchecking 'match timeline settings' and matching the resolution, but same results.

C

[Dwaine Maggart]

Windows or Linux? What OS version?

Windows 10 Pro Version 1709.

I think to get realtime playback in a system like this, you need even more powerful CPUs, more cores, more RAM, and 8GPUs. I would bet a 22-proc Intel Broadwell CPU might do it.

If you can live with doing all the work in a 4K timeline and then monitor completely in HD, this is a lot easier to do. The real challenge is to be able to throw in OFX plug-ins, NR, and blur modes, and running all this in real-time is taxing for even the biggest systems. It helps to cache one of those nodes near the end, and that will help buffer the playback. Having a RAID5 SSD as a source drive might help as well, but a lot depends on the nature of the source material.

Oh, and this won't be cheap. It will be interesting to see what happens at NAB in April, particularly on the performance/cost area for systems as well as performance increases in the new release of Resolve.

[Peter Chamberlain]

Can u render 30 seconds to DNxHR at uhd and put that in a new project and test?

Dual 18 core with half res decode in red raw settings should work, as should the optimized media test.

You have something else going on.

This might not be as effective in 14.2 as in 12.5 but these tips might be worth a try.

Check "Hide UI Overlays for optimized playback" in Master Project Settings.
Make sure your Background Caching is off. (General Options)
Make sure your Timeline (Edit Page) is set with NO Icons.
All of that helps with playback for myself.

Additionally you can also (Camera Raw) Decode at Half Res Premium and use Play Quality at Half Res Good... IF you are monitoring at HD.
(Parking on a clip shifts to a better debayer -playback does not).

If that helps - remember during Output/Delivery to (Advanced Settings DO NOT use render cache or optimized media AND DO use Force debayer to highest quality.

Here's a long shot, but you might want to check it out.

I bought a Supermicro 7048gr-tr a while back and noticed something interesting. The motherboard manual on page 1-6 says this "Warning: To avoid damaging your motherboard and components, please use a power supply that supports a 24-pin, two 8-pin and and one 4-pin power connectors.
Be sure to connect the 24-pin and the 8-pin power connectors to your power supply
for adequate power delivery to your system. The 4-pin power connector is optional;
however, Supermicro recommends that this connector also be plugged in for optimal
power delivery."

This 4-pin connector is JPWR4. It was not connected when I received my 7048gr-tr, and the two built-in power supplies didn't even have the proper type of 4-pin connectors to make the connection. My power needs are small now: 2 6-core 2.4 ghz processors and a single gtx-1080, so I haven't worried about it yet. I figured I'd ask Supermicro about it or my reseller when I upgrade to more powerful stuff and see if I can use some sort of cable adapter to make the connection. But your system is much beefier. If it was delivered with JPRW4 unconnected like mine was, maybe it's a little underpowered.

Can u render 30 seconds to DNxHR at uhd and put that in a new project and test?

Dual 18 core with half res decode in red raw settings should work, as should the optimized media test.

You have something else going on.

Peter,
Sorry for the delay on responding to this
I rendered out a 30 second clip DNxHR HQ at UHD and dropped it in a new project and had no issues with playback. I have also dropped this whole project into a new project but the issue remains. I have actually had to go to 1/8th res good decode quality to get smooth consistent playback. It also says 24 fps even though I get the jittery playback at 1/2 res good and prem decode rates.

If you have any other test suggestions let me know

For a test sake, close the scopes when in Color page; does it cause fps increase?

Piotr

Chris, even in a big linux rigs, wee don't go straight from the red raw: smart cache on will automatically convert and cache the source in an intermediate fast codec (we usually use dpx either 10 or 16 bit)

@Piotr - I will test that, I haven't gotten past the Edit page for speed tests yet .

@waltervolpatto - Good to know, I am using smart cache now. It is a bit time consuming when it re-caches on every change, need to read a bit more on how to refine that with user mode if possible. I would optimize all of my media, but this project has about 30 TB of raw footage, so optimizing at a high quality intermediate would cause a space issue for me

@Mike I did check the JPWR4 connection, and it was not connected :/ I connected it but unfortunately it didn't seem to have any impact. Good to know though!

Thanks all for the input!

C

Have you tried disconnecting the Cubix expander and just using the one 1080Ti GTX card?

On the I/O front, RED cameras can only record up to 300MB/sec, which means that any storage array that can sustain 300MB/sec (really sustain it), won't be the bottleneck. The 20% GPU utilization with 3GPUs is telling, I think.

I've had an RR-X card since buying my DRAGON. What I noticed with a 2012 Mac Pro (2x 6 core Xeon) was that without RR-X, the machine was totally CPU-bound. With the RR-X CPU dropped to 20% and the system became GPU-bound. I suspect what's killing your performance is the wavelet decoding, though I'm really surprised that 2x 18 core processors cannot handle it. Which is why I suspect that somehow the external GPU cards are somehow gumming up the works.

You might try downgrading to 14.1.1. In the move to 14.2 I lost a decent amount of performance, and had to go back.

You might try downgrading to 14.1.1. In the move to 14.2 I lost a decent amount of performance, and had to go back.

I recommend more faith in BMD Developers; I bet the 14.3.1 (or whatever the new public build will be called) is going the address the playback fps issue (including that in the Color page with Scopes displayed, which I've been complaining about in another thread of mine).



Piotr