CryoCloud staging · 3D classification · November 2026
Our new 3D classification algorithm: sorting 20M particles in 6.5h
We benchmarked 17 particle stacks containing 20 million particles with CCRefine Class3D, our new GPU classification engine and compared it to Relion's. Each run used the same particles, the same starting map and the same job settings. Using 4 GPUs for the 17 jobs sequentially, our classification finished in just 6.5 h, 3.8x faster than Relion.
When used in our production environment on multiple instances, our new classification algorithm supports throughputs of half a billion particles & >300 datasets per day for less than 1,000 USD, enabling at scale analysis for AI applications.
*Median speed-up across all 17 jobs; mean: 4.3×; total time ratio: 3.6×
17
jobs, each run by both programs, from 14 datasets
20.5 M
particles classified in total
23.2 h → 6.5 h
compute time for the classification of 20M particles using just 4x GPUs.
0.24 Å
average higher resolution of the best class vs. Relion.
RELION (the original job)
CCRefine Class3D (the rerun)
All jobs back to back
The whole benchmark as one race
Imagine running every same-machine job one after another. Each bar fills at the same pace, one tick per hour of compute - our algorithm finishes the task 3.8x faster.
What went in
The test set: 20 million particles from 17 datasets, from small to large particle stacks
Our benchmark covers a wide range of particle stacks and settings: from 62,000 to 5 million particles per stack, image boxes from 44 to 320 pixels, and various search modes. “Box” is the particle image size in pixels; “pixel” is how many ångströms one pixel covers.
| Run | Dataset | Particles | Box (px) | Pixel (Å) | Box (Å) | Classes | Iterations | How poses are searched |
|---|---|---|---|---|---|---|---|---|
| Run #2 | EMPIAR-11918 | 5.09 M | 44 | 2.61 | 115 | 5 | 6 | global searchpadding 1 |
| Run #3 | EMPIAR-13879 | 4.31 M | 96 | 2.37 | 228 | 5 | 6 | global search |
| Run #6 | EMPIAR-11134 | 1.98 M | 170 | 1.29 | 220 | 5 | 20 | global searchfast subsetspadding 1 |
| Run #4 | EMPIAR-11793 | 1.84 M | 72 | 2.53 | 182 | 5 | 6 | global search |
| Run #16 | EMPIAR-11918 | 1.34 M | 112 | 1.02 | 115 | 5 | 25 | no alignmentfast subsets |
| Run #5 | EMPIAR-10252 | 1.29 M | 192 | 1.12 | 215 | 5 | 6 | global search |
| Run #14 | EMPIAR-10342 | 758 k | 320 | 1.10 | 352 | 5 | 18 | local searchfast subsetsstops when convergedpadding 1 |
| Run #13 | EMPIAR-11316 | 662 k | 224 | 1.22 | 273 | 5 | 25 | local searchfast subsets |
| Run #17 | EMPIAR-10288 | 549 k | 140 | 1.60 | 224 | 3 | 25 | global searchfast subsetscoarser sampling |
| Run #7 | EMPIAR-10288 | 528 k | 256 | 1.29 | 330 | 5 | 25 | global searchfast subsets |
| Run K1 | EMPIAR-11442 | 476 k | 208 | 1.32 | 275 | 1 | 25 | global searchfast subsets |
| Run #15 | EMPIAR-11807 | 461 k | 168 | 1.02 | 171 | 5 | 25 | no alignment |
| Run #9 | EMPIAR-11442 | 429 k | 208 | 1.32 | 275 | 5 | 25 | global searchfast subsets |
| Run #8 | EMPIAR-13252 | 407 k | 220 | 2.49 | 548 | 5 | 23 | global searchfast subsetsstops when converged |
| Run #12 | Tilted test set | 244 k | 180 | 1.51 | 272 | 3 | 25 | global searchfast subsetscoarser samplingC3 symmetry |
| Run #11 | EMPIAR-10673 | 82 k | 120 | 2.43 | 292 | 9 | 6 | global search |
| Run #10 | EMPIAR-10250 | 62 k | 100 | 1.23 | 123 | 5 | 20 | global searchfast subsets |
Speed
Our 3D classification job is faster on every single dataset
Job time from start to finish, as the platform records it. Two jobs run the classification without alignment, which RELION runs on the CPU, which are shown in the bottom table.
Classification without alignmentRELION runs this mode on the CPU, CCRefine on the GPUs
Quality
Our 3D classification improves the resolution of the best class by 0.24 Å on average
Resolution of the best class in each job, as each program estimates it. Smaller numbers and dots shifting to the right mean higher resolution.
Right column: CCRefine minus RELION. Green means CCRefine's best class is finer.
Look for yourself
Class by class
Pick a dataset. Each row pairs a RELION class with the CCRefine class that holds most of the same particles. The pictures are flat projections of the 3D class maps. “Maps agree to” is the detail level down to which the two maps match (FSC 0.5); for scale, the next line says how alike RELION’s own classes are to each other.
Run #3 · EMPIAR-13879
Particle by particle
Do particles land in the same class?
3D classification starts from a random split, so two runs never sort every particle the same way. This also applies for RELION run twice on the same data. The chart shows the share of particles that sit in matching classes in both runs. The gray rows are RELION against itself, for scale.
The distributions from our
Particles in matching classeshigher is closer agreement
Read before quoting
Where the two differ
How much should two runs agree?
Class3D starts from a random split, so even two RELION runs on the same data put many particles in different classes: 29–56% match in our three RELION-vs-RELION repeats. CCRefine vs RELION lands in the same range (24–80%), so the disagreement is the normal run-to-run spread, not a CCRefine effect.
Run #6 · EMPIAR-11134
Only 24% of particles match, close to chance for 5 classes. RELION’s own five classes are near-copies of each other (they differ only beyond 16–18 Å), so any split between them is arbitrary. Both programs reach the same best resolution, 5.94 Å.
Runs #10 and #12
The two programs split these datasets differently (33% and 51% same class). Both splits are stable; CCRefine’s best class is finer in both (5.85 → 5.59 Å, 5.45 → 4.70 Å).
Run #16 · EMPIAR-11918
RELION left 3 of its 5 classes empty; CCRefine filled all 5. CCRefine’s best class, 2.21 Å, is close to the 2.05 Å limit set by the pixel size.
Runs #14 and #15
These are the two runs where CCRefine’s best class is a little worse: 5.50 → 5.87 Å (#14) and 2.34 → 2.37 Å (#15).
Runs #15 and #16: classification without alignment
Both jobs classify particles without re-aligning them. RELION runs this mode on the CPU (only Blush uses a GPU), so on the same 4× L4 machine its GPUs sat mostly idle, while CCRefine runs it on the GPUs. That is how the two programs behave in this mode, and these speed-ups (7.7× and 3.9×) are real for users, but they are not a GPU-against-GPU comparison,
About the resolution numbers
Each resolution is the program’s own per-class estimate from its output files, not an independent gold-standard FSC. They show the trend, a final judgement needs a separate refinement of the chosen class.
