Ryzen vs. Xeon.

[YoyoBoy]

I have a Ryzen 7 1700 PC, and something weird happened that has me wondering if there are any VRay functions that Ryzens do much faster than Xeon chips.

A test scene I have right now takes about 30 seconds per frame on my system. With 900 frames, and a Cinebench score of 1379, the scene should have cost only 7 Euros to render on Ranch Computing's farm -which is loaded with Dual Xeon E5-2697A V4s. But twice the job ended up costing ten times that, and the frames were averaging 1:17.

Any ideas what might be going on?

Thanks.

[vlado]

I'm not aware of anything specifically that Ryzen CPUs do very much better, except that they have less cores but at higher frequencies. You can get us a scene so that we can test it here on similar hardware, and maybe then we can contact Ranch Computing and ask them to look into it.

Best regards,
Vlado

[LudvikKoutny]

I have a Ryzen 7 1700 PC, and something weird happened that has me wondering if there are any VRay functions that Ryzens do much faster than Xeon chips.

A test scene I have right now takes about 30 seconds per frame on my system. With 900 frames, and a Cinebench score of 1379, the scene should have cost only 7 Euros to render on Ranch Computing's farm -which is loaded with Dual Xeon E5-2697A V4s. But twice the job ended up costing ten times that, and the frames were averaging 1:17.

Any ideas what might be going on?

Thanks.

They will charge you also the time it takes to start up 3ds Max session, load and parse the scene. They have really, really fast Xeon machines, but many of the things in the startup process do not utilize CPUs to the max. Even in relatively simple scene, the 3ds Max startup process, parsing, caching of the maps and geometry, and translating the scene can take ~ 3 minutes. That's already 6 times more than 30 seconds.

The thing is that as long as one of their super high end Xeons is reserved for your renderfarm session, it can not render others' people jobs, so even if the dual Xeon machine spends 3 minutes with just one of let's say 128 threads active at 10%, they have to charge you the full usage of that rendernode. It does not matter if the rendernode is actually used at its full speed, they just take GHz/hour performance formula, and divide or multiply it by the time that particular Xeon node has spent catering your job.

If you have very, very quick rendertimes, that are much faster than it takes to get that very slow, clumsy 3ds Max machinery up and running, then you should use their Legacy farm. They specifically mention it in the guide. The legacy farm is composed of much slower nodes, so you will pay a lot less for time wasted starting up Max, and parsing the scene.

I found it unfair at first too, but after kind of thinking about it for a while, it's really the only way to go about it. I mean imagine having your own renderfarm, and someone inexperienced sending you scenes, that are set up in a way they render only for a few seconds but take 10 minutes to translate, due to for example something like high resolution blobmesh or very complex pFlow simulation. You would have those Xeon machines sitting there doing nothing, but unable to render others' people jobs, and you would get only like 1-5% of what their time is worth.

[Neilg]

Last time we used a render farm that wasn't amazon ec2, we ran a test frame and got a time of about 30 minutes. then we send the whole scene over that was 5gb and pulled massive textures, and got some wildly inconsistent times varying from an hour to 3 hours. very awkward and difficult experience. that's not the only time thats happened too. I can only assume most people who run render farms dont have the network bandwidth to cope with the number of machines they are running.
We have a permanent ec2 now and it's the best.

[tecsun]

Neilg Many times it is not due to slow network bandwidth but a less responsive storage. Network is cheap nowadays hence it wouldn't be something that people don't want to invest in. Our render farm did suffer from file loading time before we switched from scale-up storage to scale-out type. Scale-out storage is the way to go.

How's ec2? I tried it once and feel the fee is very unexpected, probably I do not really understand their calculator well enough. I've our render farm deployed on Azure now and it works pretty well. The cost estimation is much more accurate to me. I can spin up 500+ cores with my account but still gonna be very careful with the cost.

[Miro]

YoyoBoy We tested your scene on a couple of different machines here and these are the times we got.
Generally it was slower later in the sequence, where the camera is zoomed in at the object with the glossy reflection, but we didn't even hit as much as 1 minute/frame.

Xeon е5-2699 v4 start of the sequence was at ~1.5s/frame and later on -> ~13s/frame
Xeon L5520 start at ~20s/frame; later ~32s/frame
Xeon E5520 start at ~22s/frame; later ~35s/frame
Core i7 start at ~7s/frame; later ~44s/frame
Ryzen 7 1800X ~6-7s/frame

We'll contact the render farm to look into this.

[YoyoBoy]

Fantastic! Thank you!