Minimum system specifications for standalone operation?

[bwllc]

I've got two thirds of the workflow described in the RFDiffusion paper from David Baker's lab up and running on my home PC. I've deployed RFDiffusion itself, and ProteinMPNN. I would like to add AlphaFold.

I have concerns about whether AlphaFold will run at all on my PC. I have read some crash reports, so I am looking through the AlphaFold documents for a minimum system specification. I haven't found one.

The closest thing I found in the documentation shows one working cloud PC configuration, and some performance data. Using a rather formidable looking virtual PC instance at Google Cloud, some fairly impressive inference speeds are achieved.

I would be happy to run 50 times slower than the rates shown, but I want to be certain that I can submit a 300-residue protein without causing a system crash.

What are the "must haves" for AlphaFold operation? I'll be happy to purchase a 4 TB SSD. I can't afford an A100 GPU. I don't know exactly how to compare the cloud computing specs to the specs for a standalone machine.

Thanks for any information you can provide.

[tcoates5]

I don't know about actual minimum requirements, but I do know that the hardware in a p3.2xlarge AWS instance had no such issues for proteins much larger than that. The most common hardware related cause of crashes I have seen on this repository is by far and away GPU memory, so I would say if you have a dedicated GPU with 16+GB you should be fine, and for proteins of the size you are talking about even 8 GB is likely adequate. The closest thing to minimum requirements I have seen is in the README, where it talks about the reduced_dbs version being designed for 8 CPU cores and 8 GB of RAM. I would expect headaches if you had less than 8 GB RAM for either the GPU or CPUs, and a massive speed loss if you did not have a dedicated GPU.

[tcoates5]

Also, you probably want to make sure you are using an NVIDIA GPU

[bwllc]

Thanks for your replies, @tcoates5.

I've done a fair amount of machine learning work. At this point, I think that even a casual machine learning user won't try to make things work without an NVidia GPU and CUDA.

I have an NVidia 1660 GTX Super with 6 GB of video RAM. This GPU is being used by RFDiffusion and ProteinMPNN, but you're suggesting that the RAM might not be adequate for AlphaFold. I do want the option to use the full AF database, so I might have to upgrade the GPU. If there's a way to add RAM to an existing GPU card, I am unaware of it. The 4060 Ti with 16 GB RAM is at the upper end of my upgrade budget, but it's nowhere near as expensive as an A100.

I hope some other people can share their experiences as well.

[tcoates5]

Happy to help! I would recommend trying it out as soon as you have adequate SSD storage, I looked back at some other runs I did and it looks like I have run proteins double that size on 8GB video cards without crashes.