I've always thought of enterprise grade GPUs as overkill (leftover thinking from buying 8-GPU workstations in 2018), but there are a significant number of tasks on cryoSPARC that are single GPU, so I guess I probably need to change my thinking...

Thanks for all of the details! Do you have a post detailing your lab's computational set-up? I need to get a new rig, obviously. If not, I'd appreciate to hear what you use (can be basic like "everyone has a workstation with 2xA6000, 2x20 core Intel CPU, and 256 GB RAM".

We are using 3 classes, 5->2.3 Ang initial and high res cutoff, 0.005 Ang fourier radius step size. 300 and 1000 initial and final minibatch size, center structures in real space off.

We can see all secondary structure and easily dock a model, but there are still many artifacts (random spikes, etc.). At this rate it's about 10 days more computation to finish (lol we need a new computer). Any insight into when in the HR-HAIR you started seeing high quality maps?

Hi Oli, awesome work! HR-HAIR is definitely giving us interpretable volumes. However, we are on old 2080 GPUs, so the ab initio is very slow. In 3 days we are at iteration 3800 out of 11,000 for ~150k particles. I'm wondering if you saw considerable increase in map quality in the later iterations?