I’m seeking feedback from the V-Ray committee about V-Ray performance across different CPU setups and render nodes. We are transitioning our render-farm from older BOXX blades on-premises to a cloud solution, aiming to move 40 Windows systems and a Deadline manager to equivalent cloud-based VMs for improved speed and reliability.
To identify the best VM instance for performance and cost, I tested typical render jobs using five cloud instances and our legacy node. Tests involved three past projects and two custom 3ds Max scenes: one with Anima characters and animated materials, and another with iToosoft trees and Rail Clone buildings. Frame render times ranged from 2–60 minutes for old projects, 1.5–8.5 minutes for the Anima scene, and 45–90 minutes for the iToosoft scene.
Unexpectedly, newer AMD-based instances sometimes performed worse than the old BOXX (dual Intel Xeon) system. While some results matched expectations, others did not; for example, the newest instance was occasionally four times faster, but in other cases, the legacy system beat the cloud VMs. I understand VM overhead exists, but these discrepancies seem too large.
Can anyone explain what factors might cause such variations? Has anyone else observed inconsistent rendering performance between cloud/data center systems or workstations with differing specs? I have been consulting the cloud vendor regarding these issues, but I wanted to hear from you all and from Chaos to see what you all might think about what I have experienced.
Your older system is probably better at single threaded setup tasks. Generally AMD out performs in render tasks, but in tests with AWS nodes the Intel and AMD were so close to the same that is wasn’t worth restricting them to one or the other. Given that availability also varied it was more flexible to use both.
When using VMs there is also the VM setup time involved. If you are using Deadline with AWS it must restart a windows machine the first time the machine is brought online (in order to change the machine name). This makes the initial frame from a node take a while. If you use bare metal instances this takes a VERY long time (bare metal can take many minutes to reboot-- sometimes close to 15mins). So bare metal is a no-no.
If you have the time more, slower nodes is often better/cheaper than fewer faster nodes because so much of the setup is single threaded and does not benefit from the high core VMs. x24 seem like a decent sweet spot, x32 is good if you have slower frames, x48 only if you have really slow frames. You can try running multiple instances of the render on each node, which can help with the single threaded issue, but we had less luck with this than hoped. Nothing aws offers reaches the performance of a local 7995wx, though the x48 comes close.
Are you using Deadline with aws for your cloud, or some other setup?
I am running legacy Deadline on an internal server and not the cloud variant. As for performance, do you really think an 8-year-old CPU would have better performance than a current CPU, especially with thread scheduling? As part of the benchmarking, I am looking at several hundred frames, to allow for each test system to ease into anything. In fact, the legacy system is at a disadvantage in the tests as it must pull all the render data over the WAN from the cloud datacenter to get started, though this would go away once the job is loaded and the first frame has rendered. Really what I need to do is run tests against some modern workstations running high core-count CPUs (AMD or Intel) to observe the results. It might make better sense to just run workstation systems in a local data center. The costs would be about the same, if not a little cheaper, even with the data center costs and CAPEX hardware.
Just to add a data point from a different angle: we run a small render farm with V-Ray 7 on Linux (RHEL-based), rendering via .vrscene archives exported from 3ds Max + V-Ray on Windows. Same hardware, Linux Standalone consistently gives us ~16% better VRayBenchmark scores compared to Windows.
Our workflow: 3ds Max + V-Ray → export .vrscene archive → render on V-Ray Standalone for Linux (or via AppSDK DR2). Very stable, no licensing overhead per node, and the performance gain is a nice bonus.
Bonus: no Cortana, no ads, no background telemetry — a full render node runs at ~800 MB RAM idle. Clean and fast.
For anyone with mixed environments, it’s worth considering.
Rock solid — with V-Ray Standalone I have never had a DR node lose something and render a bucket slightly differently than the others.
And the performance gap vs. Windows grows with core count — the more cores your node has, the more Linux pulls ahead.
I have not done any performance testing in that realm in a very long time and then it provided no benefit. If anything, it was a disadvantage since vrscene was not particularly robust and it dropped much. We run enterprise Windows and Deadline as a service, so I suspect there is really no overhead there either. I concede that perhaps there might be a slight advantage these days, but 3ds max and V-Ray are not the only thing we run on these render nodes.