CryoCloud staging · 3D classification · November 2026

Our new 3D classification algorithm: sorting 20M particles in 6.5h

3.9× faster*

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.

RELION
23 h 12 min
CCRefine
6 h 31 min
CCRefine finished 16 h 41 min earlier

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.

RunDatasetParticlesBox (px)Pixel (Å)Box (Å)ClassesIterationsHow poses are searched
Run #2EMPIAR-119185.09 M442.6111556global searchpadding 1
Run #3EMPIAR-138794.31 M962.3722856global search
Run #6EMPIAR-111341.98 M1701.29220520global searchfast subsetspadding 1
Run #4EMPIAR-117931.84 M722.5318256global search
Run #16EMPIAR-119181.34 M1121.02115525no alignmentfast subsets
Run #5EMPIAR-102521.29 M1921.1221556global search
Run #14EMPIAR-10342758 k3201.10352518local searchfast subsetsstops when convergedpadding 1
Run #13EMPIAR-11316662 k2241.22273525local searchfast subsets
Run #17EMPIAR-10288549 k1401.60224325global searchfast subsetscoarser sampling
Run #7EMPIAR-10288528 k2561.29330525global searchfast subsets
Run K1EMPIAR-11442476 k2081.32275125global searchfast subsets
Run #15EMPIAR-11807461 k1681.02171525no alignment
Run #9EMPIAR-11442429 k2081.32275525global searchfast subsets
Run #8EMPIAR-13252407 k2202.49548523global searchfast subsetsstops when converged
Run #12Tilted test set244 k1801.51272325global searchfast subsetscoarser samplingC3 symmetry
Run #11EMPIAR-1067382 k1202.4329296global search
Run #10EMPIAR-1025062 k1001.23123520global 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.

RELIONCCRefine
DatasetRuntimesSpeedup
Run #3 · EMPIAR-138794.31 M particles · box 96 · 5 classes
3 h 27 min
42 min
4.9×
Run #2 · EMPIAR-119185.09 M particles · box 44 · 5 classes
2 h 31 min
42 min
3.6×
Run #13 · EMPIAR-11316662 k particles · box 224 · 5 classes
2 h 29 min
1 h 31 min
1.6×
Run #4 · EMPIAR-117931.84 M particles · box 72 · 5 classes
1 h 40 min
18 min
5.6×
Run #6 · EMPIAR-111341.98 M particles · box 170 · 5 classes
1 h 38 min
42 min
2.4×
Run #5 · EMPIAR-102521.29 M particles · box 192 · 5 classes
1 h 32 min
15 min
6.1×
Run #7 · EMPIAR-10288528 k particles · box 256 · 5 classes
1 h 26 min
23 min
3.8×
Run #14 · EMPIAR-10342758 k particles · box 320 · 5 classes
1 h 11 min
20 min
3.5×
Run #9 · EMPIAR-11442429 k particles · box 208 · 5 classes
54 min
13 min
4.1×
Run #8 · EMPIAR-13252407 k particles · box 220 · 5 classes
47 min
18 min
2.6×
Run #17 · EMPIAR-10288549 k particles · box 140 · 3 classes
41 min
11 min
3.8×
Run K1 · EMPIAR-11442476 k particles · box 208 · 1 classes
37 min
8 min
4.6×
Run #11 · EMPIAR-1067382 k particles · box 120 · 9 classes
31 min
4 min
7.5×
Run #12 · Tilted test set244 k particles · box 180 · 3 classes
24 min
8 min
2.9×
Run #10 · EMPIAR-1025062 k particles · box 100 · 5 classes
13 min
3 min
5.0×

Classification without alignmentRELION runs this mode on the CPU, CCRefine on the GPUs

Run #15 · EMPIAR-11807461 k particles · box 168 · 5 classes
2 h 07 min
17 min
7.7×
Run #16 · EMPIAR-119181.34 M particles · box 112 · 5 classes
1 h 04 min
16 min
3.9×

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.

RELIONCCRefine
12 Å10 Å8 Å6 Å4 Å2 Å← coarserfiner detail →Run #15 EMPIAR-11807CCRefine Class3D benchmark | CryoCloudCCRefine Class3D benchmark | CryoCloud+0.03 ÅRun #16 EMPIAR-11918CCRefine Class3D benchmark | CryoCloudCCRefine Class3D benchmark | CryoCloud−0.46 ÅRun #13 EMPIAR-11316CCRefine Class3D benchmark | CryoCloudCCRefine Class3D benchmark | CryoCloudsameRun #17 EMPIAR-10288CCRefine Class3D benchmark | CryoCloudCCRefine Class3D benchmark | CryoCloud−0.19 ÅRun #7 EMPIAR-10288CCRefine Class3D benchmark | CryoCloudCCRefine Class3D benchmark | CryoCloud−0.35 ÅRun #3 EMPIAR-13879CCRefine Class3D benchmark | CryoCloudCCRefine Class3D benchmark | CryoCloud−0.21 ÅRun #4 EMPIAR-11793CCRefine Class3D benchmark | CryoCloudCCRefine Class3D benchmark | CryoCloudsameRun #2 EMPIAR-11918CCRefine Class3D benchmark | CryoCloudCCRefine Class3D benchmark | CryoCloudsameRun #12 Tilted test setCCRefine Class3D benchmark | CryoCloudCCRefine Class3D benchmark | CryoCloud−0.75 ÅRun #14 EMPIAR-10342CCRefine Class3D benchmark | CryoCloudCCRefine Class3D benchmark | CryoCloud+0.37 ÅRun K1 EMPIAR-11442CCRefine Class3D benchmark | CryoCloudCCRefine Class3D benchmark | CryoCloud−0.44 ÅRun #10 EMPIAR-10250CCRefine Class3D benchmark | CryoCloudCCRefine Class3D benchmark | CryoCloud−0.27 ÅRun #6 EMPIAR-11134CCRefine Class3D benchmark | CryoCloudCCRefine Class3D benchmark | CryoCloudsameRun #5 EMPIAR-10252CCRefine Class3D benchmark | CryoCloudCCRefine Class3D benchmark | CryoCloud−0.17 ÅRun #8 EMPIAR-13252CCRefine Class3D benchmark | CryoCloudCCRefine Class3D benchmark | CryoCloud−1.20 ÅRun #9 EMPIAR-11442CCRefine Class3D benchmark | CryoCloudCCRefine Class3D benchmark | CryoCloud−0.45 ÅRun #11 EMPIAR-10673CCRefine Class3D benchmark | CryoCloudCCRefine Class3D benchmark | CryoCloudsame

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

4.31 M particlesbox 96 px at 2.37 Å5 classes, 6 iterationsglobal searchtime 3 h 27 min → 42 minsame class: 70% of particles
RELION 5
CCRefine 5
Share of particles35.2% · 33.8%
Resolution5.99 Å · 5.69 Å
Maps agree to≤ 4.7 ÅRELION's closest other class: 56.9 Å
RELION 4
CCRefine 4
Share of particles26.1% · 24.7%
Resolution7.58 Å · 7.34 Å
Maps agree to≤ 4.7 ÅRELION's closest other class: 56.9 Å
RELION 1
CCRefine 1
Share of particles23.1% · 23.5%
Resolution5.06 Å · 4.84 Å
Maps agree to≤ 4.7 ÅRELION's closest other class: 56.9 Å
RELION 2
CCRefine 2
Share of particles8.2% · 9.5%
Resolution9.10 Å · 8.75 Å
Maps agree to≤ 4.7 ÅRELION's closest other class: 56.9 Å
RELION 3
CCRefine 3
Share of particles7.4% · 8.5%
Resolution9.10 Å · 9.10 Å
Maps agree to≤ 4.7 ÅRELION's closest other class: 56.9 Å

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

CCRefine vs RELIONRELION vs RELION repeat

Particles in matching classeshigher is closer agreement

RELION vs RELIONRun #17 data (pad 1 vs 2)
56%
RELION vs RELION857 k particles, identical setup
46%
RELION vs RELION413 k particles, same setup apart from one GPU speed option
29%
CCRefine vs RELION repeatRun #17 data, against the second RELION run
51%
Run #2 · EMPIAR-119185 classes
80%
Run #13 · EMPIAR-113165 classes
71%
Run #3 · EMPIAR-138795 classes
70%
Run #5 · EMPIAR-102525 classes
68%
Run #4 · EMPIAR-117935 classes
67%
Run #15 · EMPIAR-118075 classes
62%
Run #16 · EMPIAR-119185 classes
57%
Run #8 · EMPIAR-132525 classes
56%
Run #17 · EMPIAR-102883 classes
55%
Run #11 · EMPIAR-106739 classes
55%
Run #14 · EMPIAR-103425 classes
52%
Run #12 · Tilted test set3 classes
51%
Run #7 · EMPIAR-102885 classes
42%
Run #10 · EMPIAR-102505 classes
33%
Run #9 · EMPIAR-114425 classes
29%
Run #6 · EMPIAR-111345 classes
24%

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.

How we measured

Each CCRefine job was cloned from its RELION job: with the same same parent particles, reference, mask and setup arguments, Particle counts, box and pixel size, class counts and iterations are read from the data files each job wrote; search modes and GPU use from each RELION job’s command. Times are the job’s running-to-finished span from the platform database. Class shares and resolutions are the values each job reports for its last iteration. Particle matching pairs classes so the most particles agree (Hungarian matching on the class-by-class count table), and map agreement is the FSC 0.5 crossing between the paired maps after whole-pixel alignment.

Frequently Asked Questions

If it's not covered here, please reach out to us.
…or start your free trial and see for yourself.

Is there a free trial?

Yes — when you sign up, you'll get instant access to a 30-day free trial with full access to the CryoCloud web application. You'll receive 20 hours of compute time and 2 TB of storage. No setup or infrastructure needed. After the trial, you can upgrade to a paid subscription or contact us for custom plans — no interruption to your data.

How can I start using CryoCloud?

Either go directly to out signup page or simply send us an email to hi@cryocloud.io so we can invite you to our application. You can find more information on how to get started here.

Do I need to install anything?

No - you really just need a (simple) computer, internet access and an internet browser. We took care of everything else. Read more on it here

How does your risk-free trial work?

If you are unsatisfied, you can cancel your plan within 30 days and we will refund you at 100%. You can deactivate the subscription from your account settings.

Frequently Asked Questions

If it's not covered here, please reach out to us.
…or start your free trial and see for yourself.

Is there a free trial?

Yes — when you sign up, you'll get instant access to a 30-day free trial with full access to the CryoCloud web application. You'll receive 20 hours of compute time and 2 TB of storage. No setup or infrastructure needed. After the trial, you can upgrade to a paid subscription or contact us for custom plans — no interruption to your data.

How can I start using CryoCloud?

Either go directly to out signup page or simply send us an email to hi@cryocloud.io so we can invite you to our application. You can find more information on how to get started here.

Do I need to install anything?

No - you really just need a (simple) computer, internet access and an internet browser. We took care of everything else. Read more on it here

How does your risk-free trial work?

If you are unsatisfied, you can cancel your plan within 30 days and we will refund you at 100%. You can deactivate the subscription from your account settings.

Frequently Asked Questions

If it's not covered here, please reach out to us.
…or start your free trial and see for yourself.

Is there a free trial?

Yes — when you sign up, you'll get instant access to a 30-day free trial with full access to the CryoCloud web application. You'll receive 20 hours of compute time and 2 TB of storage. No setup or infrastructure needed. After the trial, you can upgrade to a paid subscription or contact us for custom plans — no interruption to your data.

How can I start using CryoCloud?

Either go directly to out signup page or simply send us an email to hi@cryocloud.io so we can invite you to our application. You can find more information on how to get started here.

Do I need to install anything?

No - you really just need a (simple) computer, internet access and an internet browser. We took care of everything else. Read more on it here

How does your risk-free trial work?

If you are unsatisfied, you can cancel your plan within 30 days and we will refund you at 100%. You can deactivate the subscription from your account settings.

Ready to skip the queue?
Start analyzing your cryo-EM data today.

We would love hear about your cryo-EM goals and discuss how we can accelerate them.

  • No infrastructure needed

  • Streamlined onboarding & legal support

  • Highest security

Ready to skip the queue?
Start analyzing your cryo-EM data today.

We would love hear about your cryo-EM goals and discuss how we can accelerate them.

  • No infrastructure needed

  • Streamlined onboarding & legal support

  • Highest security

Ready to skip the queue?
Start analyzing your cryo-EM data today.

We would love hear about your cryo-EM goals and discuss how we can accelerate them.

  • No infrastructure needed

  • Streamlined onboarding & legal support

  • Highest security

Accelerate scientific discovery with our cloud-native cryo-EM platform, from data collection to high-resolution structure determination.

Subscribe to our newsletter

© 2021-2026 CryoCloud B.V. All rights reserved.

Accelerate scientific discovery with our cloud-native cryo-EM platform, from data collection to high-resolution structure determination.

Subscribe to our newsletter

© 2021-2026 CryoCloud B.V. All rights reserved.

Accelerate scientific discovery with our cloud-native cryo-EM platform, from data collection to high-resolution structure determination.

Subscribe to our newsletter

© 2021-2026 CryoCloud B.V. All rights reserved.