Budget (including currency): High/Don't care how much
Country: Poland
Games, programs or workloads that it will be used for: Training Residual Neural Networks, Analyzing videos in 2D, reconstructing 2D into 3D (DeepLabCut, Anipose), probably spike sorting (electrophysiology analysis) and possibly synced recording from 6-8 cameras in at least 30 fps (around 720p)
Other details (existing parts lists, whether any peripherals are needed, what you're upgrading from, when you're going to buy, what resolution and refresh rate you want to play at, etc): We have a setup with RTX 2080Ti and analyzing our data becomes very slow (we have 3 synced cameras, going for 6-8 in the future - synced during recording, primary camera used as trigger for secondary cameras).
Hello,
I'm a neuroscientist specifically working with 3D pose estimation and electrophysiological recordings in socially interacting animals. So our current setup for neural net training and video analysis is based around single RTX 2080Ti, which was fine for 2D data, single movie per animal etc. Now since we started experiments with multiple cameras and 3D reconstruction it has become painfully slow (like a week of constant analysis to get our data from all the movies from a single experiment).
Though I know my way around computers, I have never build anything so specifc and complex. What I was thinking about was a quad sli of Titans RTX (for that much needed VRAM), AMD Ryzen Threadripper 3990X (is this a good choice?) and than problems start. As you may have guessed we don't care much for rgb and fancy looks.
1. What we need is a motherboard able to support 4 x Titan RTX obviously but we also might need an expansion card for usb ports (we use usb cameras, and each of them needs the whole bandwith of usb 3.0).
2. Is ECC RAM worth it and does it make sense for our use? Is it better to go for lower latency or higher clock speed?
3. With so much data coming from all the cameras, NVMe M2 SSDs are a must, plus some HDD(s) to put the data after whole day of recording. Since we can go for recording with compression we can use less disk writing capabilites, but I have no idea how cameras actually utilize cpu. Is it cores/threads/what heavy?
4. Since neural nets training and analyzing data using them is gpu heavy workload lasting hours upon hours (hopefully with the new setup, not days like it's now) they need exceptional cooling and that gets us to the last thing.
5. Cooling in general and dust-proofing. It's going to get hot in there and we cannot have the computer in a sterile/dustless environment, there will be animals, cage bedding, all this stuff flying around. (Though experimental rooms are temp controlled and it's around 22 degrees celsius)
Every bit of help would be great, I'm mostly worried about the part with video recording. I know more about neural nets and how they utilize gpu than I know about what multiple cameras need to work in a synced manner for a longer period of time.
