Render Benchmark Progress

I’ve been working steadily to improve rendering performance. It’s good, but it would be great to show 60 fps for molecular dynamics even of large proteins.

So I merged in a render benchmark (created with Claude Code) and have been working steadily on a few things since 2.1. (Claude has been planning / suggesting since my knowledge of rendering pipelines is limited.)

There are two key metrics:

  • camera performance (i.e., just rotating the model around)
  • “rebuild” performance (i.e., updating coordinates to rebuild the scene)

All times here are on my stack MacBook M4 - decent performance, but no dedicated GPU. (I’d welcome some help testing on other systems.)

Cartoons

System Atoms Rebuild, 2.1 Rebuild, now Camera-only, 2.1 Camera-only, now GPU memory uploaded
4HHB (hemoglobin) 4,779 106 fps 214 fps 945 fps 984 fps 14.5 → 9.6 MB
6VXX (SARS-CoV-2 spike) 23,694 19 fps ~155 fps 555 fps 595 fps 72 → 49 MB
2MS2 (MS2 capsid) 183,900 2.6 fps 6.1 fps 108 fps 179 fps 584 → 250 MB
9MB0 (1M-atom assembly) 992,820 0.5 fps ~10 fps 23 fps 120 fps 3.3 GB → 347 MB

In short, camera-only is plenty fast now, but we still have some work on cartoons. I’d love to get all of these to over 15-20 fps for the 2.2 release, or better yet 30 fps.

Ball-and-stick

System Atoms Rebuild, 2.1 Rebuild, now Speed-up
Water box 100,000 25 ms 12 ms 2.1×
2MS2 183,900 73 ms 39 ms 1.9×
9MB0 992,820 433 ms 220 ms 2.0×

So that’s ~25 fps for ~200k atoms under dynamics. I’m hoping we can get to 60 fps for 100,000 atoms and at least 30 fps for 250k atoms for the 2.2 release.

It would be great to get the 9MB0 system to ~10-15 fps for 2.2 as well (so another 2-3x performance gain.)

In other words, render performance is definitely better - though there’s still some work to do.

  • parallel CPU or GPU tessellation of cartoons and cylinders
  • parallel evaluation of electrostatic potential (just landed - really fast)
  • adapting these gains to other render styles (VdW, licorice, wireframe, etc.)
1 Like