What GPU params are important for acceleration in Fusion?

Hi everyone,
I'm currently piecing together a new machine and pondering the question of GPU. What GPU parameters are most relevant when it comes to acceleration for Fusion? Which types of cores (CUDA cores, RT cores, Tensor cores) are used by Fusion?

Would appreciate any insight!
J.

In Fusion Studio it would be single card RAM. I use 3 1080tis 11Gb each and will have to cache particular node trees. The multi GPU use can be decent but because your main one also handles your screen, it will eat up the VRAM a little quicker.

It actually runs better for me in Resolve. It will pull the other two GPUs and distribute a little more evenly depending on what I'm doing.

The Rtx3090s are looking real good right now because of the 24Gb of VRAM. Cough, Redshift, cough.

In Resolve, the VRAM situation is more important. Color grading relies on it so coloring 4k footage gets painful on less than 8GB of VRAM. I can hit walls after 15 nodes or so when on my setup at 1080p.

If you haven't had a chance, go to Puget Systems and check out their card bench marks. They run Resolve and Fusion tests for just this type of question.

Hey, thx for reply and for the tipp the Puget Systems benchmarks. In the Resolve benchmark there one 3090 nearly reaches a dual 2080Ti config - wow!

As I am going to use the card also for Redshift (and also for simming in Houdini), lots of VRAM are a good thing for sure. And for rendering, splitting the VRAM to several cards is not helpful I guess.

I was also thinking about an A6000 (or similar workstation card) - they have lots of VRAM. But as I want to dive into realtime graphics (Unreal) these cards are not the best choice (too few RT cores).

Well, seeing as your workflow looks a lot like mine, I'd say spend your money on more cards instead of the A6K. They are twice as expensive and only as fast as a 3090. Since you use Redshift, it has out of card memory so 24GB won't be a real ceiling where it matters. I've never had an issue with maxing VRAM even on my cards. Rendering one high grade frame with EXR AOVs takes at least minute so going out of memory doesn't really slow down my renders. Then again I don't run environmental forest renders or anything like that. Any major instancing is done with rsProxies so 24GB should be just fine. Getting 2 RTX 3090s should be good for at least 4 years. I think the use of GPU for everything is awesome but I can actually see where we're hitting ceilings on what we want to be GPU vs nonGPU just for memory management. Crazy.

To you original question of building a new machine... The most under rated thing for Fusion is drive I/O. Widen that as much as possible. Like 4 NVME RAID 0 wide. You will never go wrong with faster access to your footage. GPU doesn't do much if you have to wait for file access. It's the slowest part of my existing workflow.

I'm only planning to use Redshift - haven't so far. Total virgin re GPU rendering So, good to know about the out of card memory option in Redshift - thx!

For huge scenes I will continue to use Houdini's Mantra anyway I guess (or maybe I dare to jump into Clarisse...


Thx for the hint about storage bandwidth! I'm switching from Mac to PC, so I'm new to these options. What does the "wide" say in this context?

So on a Mac, you have thunderbolt for extra drives. It has a theoretical 25 - 40 Gbps bandwidth for your data. And EXR can be from 200mb to 5Gb per frame. If you're trying to pull 24 frames per second (ha!) You can do the math for what you would need to go real time. It works like that for any file or video.

You're working up hill in two ways with Windows. Thunderbolt is rare enough that you have to look for it if you want to use an external RAID at that bandwidth. That leaves you with internal RAIDs on a card. They can be awesome but the Widows file system isn't as fast as Mac OSX. In either case it has to work hard to get faster file access. That's why I said 4 NVMEs in Raid 0. It will give you as much bandwidth as the RAID card can interact with the PCIE Bus. There's a whole conversation about PCIE lanes and which processors to get too. Intel has less lanes so you get Thunderbolt but less bandwidth for GPUs and Raid cards. AMD has more lanes but only one board I know of has Thunderbolt?

Getting 2 RTX 3090s should be good for at least 4 years.

The second card won't be used, at least currently, and we don't have any indication from BMD when it would. The A6000 can also display 10 bit color, which the 3090 cannot.

You sure about the 10bit thing Chad? I have been displaying 10bit colour on the last 3 generations of NVidia cards (e.g. 1080, 2080ti ...) ?

Whew, made me question my sanity for a second. Went down the rabbit hole on the web and was looking at nvidia forums. It was talking about only supporting 8bit and 12 bit on consumer cards and 10 bit on Quadros/A series. That was only over HDMI 2.0. So then I went to the Nvidia control panel and checked out mah settings. Then Windows Display Settings... All good.

The second card will absolutely be used in Resolve and to amazing effect in Houdini with Redshift. I can get my three 1080tis going well in Resolve. Since Resolve 17, I've moved my GPU configuration all over. From not using my main card and only using the other two, back to using Auto. It depended on what I was doing in other software. Like you can't run After Effects and Fusion at the same time without weird things happening with OpenGL? So I'd keep Resolve/Fusion off of the main card to let Ae have it.

You sure about the 10bit thing Chad? I have been displaying 10bit colour on the last 3 generations of NVidia cards (e.g. 1080, 2080ti ...) ?

Interesting. I wonder when that changed.

The second card will absolutely be used in Resolve and to amazing effect in Houdini with Redshift.

But if you wanted to run Fusion faster, the money that was spent on the 3090 could buy more RAM or other things that will make Fusion faster.

I generally recommend RAM over GPU anyway for Fusion as the performance impact of a cache miss is orders of magnitude greater than the impact of a GPU upgrade.

Totally agree. One 3090 is equivalent to 3 1080tis. For compositing and normal color work it's good to go. 128Gb of RAM is like $400.00 so I would assume that's going to be a given.

3D rendering is counted in days so an extra 2k to half your render time is a no brainer. I often do renders with full AOVs overnight with Redshift. With VRAY or Arnold it would be impossible. Two 3090s makes rendering a lot of things like rendering out of Fusion. It's slow but useable for lighting and look dev and rough approvals with same day turn around. You can get a good 960x540 frame in under 10 seconds. In 3D that's a damn miracle.

Thanks guys for all the input!

My takeaway is:
1. If it's only about Fusion the GPU probably is less important then memory and storage bandwidth
2. Multiple cards should be usable with Fusion
3. 10bit seems to be a bit of a question mark with a Nvidia consumer cards - I've read that all over the place that using e.g. Intensity Pro 4K would be advisable if HDR is essential.

As I will also (actually mainly) do 3D in Redshift and also realtime with Unreal I guess I will still go with the consumer RTX card (hopefully 3090). And get lots of RAM and some nice NVMe M.2s.

Thx a bunch to all!
J.

A good machine could be something like this...

32Core AMD Ryzen Threadripper 3970X
128GB RAM
RTX3090 24GB

512 NVME System Drive

High Point SSD7103 Bootable 4X M.2 NVMe RAID Controller (2TB NVMEsx4) Online Storage

4TB Hard Drive for Nearline Storage

You could go big and get more nvme storage or make it into two separate volumes but, at those speeds, I don't think you need to split your cache and media drive like you would do some/most cases. I tend to actually nest my cache folders into the current project folder setup. That way when I come back to the project, I don't wait for caching again.

If your bank account can handle it, then get that second RTX3090 to speed up your... everything.

The 10bit thing is a non issue. It WAS an issue but not for a while now. Use what you want. Your monitor is going to be more expensive than the video cards so make sure to pay attention to that. DCI-P3 100% is the goal. See what your budget can handle.

Good Luck! I hope you're build is a beast!

It also depends on what you want to do. If you're doing very simple comps, then I/O is the bottleneck. If you're doing complex, but subjective comps, where you need a lot of interactivity and iteration, then RAM is the bottleneck. If you're doing very repetitive, pipelined comps, then the CPU and GPU speed is the bottleneck.

2. Multiple cards should be usable with Fusion

Should, but aren't.

3. I've read that all over the place that using e.g. Intensity Pro 4K would be advisable if HDR is essential.

There's no HDR metadata support from the Intensity Pro 4K, AFAIK. You'd need a Decklink (with HDMI 2.0a) for that.