Performance Metrics

CryoSPARC Live Performance Metrics

CryoSPARC Live is built and tuned for high performance pre-processing and streaming reconstruction of single particle data, and can use multiple concurrent GPUs for to maximize throughput.

CryoSPARC Live preprocessing includes four steps: (1) motion correction, (2) CTF estimation, (3) particle picking and (4) extraction.

CryoSPARC Live can sustain a throughput of 450 or more exposures per hour, per GPU, for K3 data. On a 4-GPU machine, that can scale to 1800+ exposures per hour! For K2 or Falcon data, performance can be even higher, upwards of 650 exposures per hour per GPU.

Depending on your hardware configuration (particularly raw data storage disk access speed), each preprocessing worker can sustain a throughput of at least one movie every 30 seconds, which is equal to ~2,500 movies per day per GPU.

In our internal tests, we have seen performance on well-tuned systems (like the testing hardware below) reaching up to 8,000 movies per GPU per day. See the Hardware Configurations used for Benchmarking section to see details on what hardware was used to run the benchmarks.

All 3D renderings were captured in ChimeraX from maps created by cryoSPARC Live.

Hardware Configurations Used for Benchmarking

All pre-processing timings were measured with Configuration 1, unless otherwise noted.

Component

Configuration 1

Configuration 2

Configuration 3

CPU

AMD Ryzen Threadripper 2950x

AMD Ryzen Threadripper 3960x

AMD Ryzen Threadripper 3960x

Memory Bandwidth

128 GB/s

144GB/s

144GB/s

RAM

128GB DDR4 2666MHz

256GB DDR4 2933MHz

256GB DDR4 2933MHz

GPU 0

Quadro GV100

Quadro RTX 8000

GeForce RTX 3090

GPU 1

Quadro GV100

Quadro RTX 8000

GeForce RTX 3090

GPU 2

Quadro RTX 5000

GTX 1080Ti

-

GPU 3

GTX 1080Ti

Tesla K40c

-

Fast CPU memory bandwidth is a major contributing factor to high performance in cryoSPARC Live. Please make note of this metric when selecting your system's CPU and RAM.

K2 MRC (HA Trimer)

Benchmark results for 668 MRC-format uncompressed movies from a GATAN K2 4k × 4k detector. The first 40 of 100 frames were used.

Exposures from this dataset were captured with the stage tilted 40º.

Particles were selected with the Template Picker strategy. Streaming 2D Classification, Ab-initio Reconstruction and Streaming Refinement yielded 3.0Å resolution map from ~230,000 particles.

Data Properties

Property

Value

Detector

Gatan K2

Number of Movies

668

File Format

MRC

Frame Size

3838 x 3710

Frames per Movie

100 (40 used)

Pixel Size

1.13Å

Particle Extraction Box Size

144 × 144

Benchmarks

Metric

Value

Movies Pre-processed Per Hour Per GPU

430

Movies Pre-processed Per Day Per GPU

10290

Average Pre-processing Time Per Movie

8.4s

K2 TIFF (Nav1.7)

Benchmark results for ~24,000 TIFF-LZW compressed movies from a GATAN K2 4k × 4k detector.

Particles were selected with the Blob Picker strategy. Streaming 2D Classification, Ab-initio Reconstruction and Streaming Refinement yielded 3.3Å resolution map from ~300,000 particles.

Data Properties

Property

Value

Detector

Gatan K2 Summit

Number of Movies

25084

File Format

TIF-LZW

Frame Size

3838 × 3710

Frames per Movie

40

Pixel Size

0.85Å

Particle Extraction Box Size

512 × 512

Particle Extraction Bin Size

256 x 256

Applied Symmetry

C2

Benchmarks

Metric

Value

Movies Pre-processed Per Hour Per GPU

650

Movies Pre-processed Per Day Per GPU

15600

Average Pre-processing Time Per Movie

5.5s

K2 super-res TIFF (T20S)

Benchmark results for ~200 TIFF-LZW compressed movies from a GATAN K2 detector with super-resolution capture.

Particles were selected with the Template Picker strategy. Streaming 2D Classification, Ab-initio Reconstruction and Streaming Refinement yielded a 2.5Å resolution map from ~130,000 particles.

The target T20S Proteasome has D7 symmetry.

Data Properties

Property

Value

Detector

Gatan K2 (super-res)

Number of Movies

196

File Format

TIFF-LZW

Frame Size

7420 × 7676

Frames per Movie

38

Pixel Size

0.6575Å

Particle Extraction Box Size

448 × 448

Applied Symmetry

D7

Benchmarks

Metric

Value

Movies Pre-processed Per Hour Per GPU

254

Movies Pre-processed Per Day Per GPU

6096

Average Pre-processing Time Per Movie

14.2s

K3 TIFF (TT-OAD2)

Benchmark results for ~200 TIFF-LZW compressed movies from a GATAN K3 detector. The first 40 of 64 frames were used.

Particles were selected with the Blob Picker strategy. Post-processing (2D Classification, Refinement, etc.) was not run on this dataset.

Exposures in this dataset were captured with beam-induced tilt.

Data Properties

Property

Value

Detector

Gatan K3

Number of Movies

3159

File Format

TIFF-LZW

Frame Size

5760 × 4092

Frames per Movie

62 (40 used)

Pixel Size

0.826Å

Particle Extraction Box Size

144 × 144

Benchmarks

Metric

Value

Movies Pre-processed Per Hour Per GPU

420

Movies Pre-processed Per Day

10050

Average Pre-processing Time Per Movie

8.6s

K3 super-res TIFF (TT-OAD2)

Benchmark results using super-resolution variants from super-resolution variant of previous dataset. Only the first 40 frames of each exposure were used.

Data Properties

Property

Value

Detector

Gatan K3 (super-res)

Number of Movies

4259

File Format

TIFF-LZW

Frame Size

11520 × 8184

Frames per Movie

67 (40 used)

Pixel Size

0.413Å

Particle Extraction Box Size

288 × 288

Benchmarks

Metric

Value

Movies Pre-processed Per Hour Per GPU

192

Movies Pre-processed Per Day Per GPU

4608

Average Pre-processing Time Per Movie

18.7s

Falcon III TIFF (PAC1)

Benchmark results for ~3000 TIFF-LZW compressed movies from a Falcon III detector.

Particles were selected with the Blob Picker strategy. Post-processing (2D Classification, Refinement, etc.) was not run on this dataset.

Data Properties

Property

Value

Detector

TFS Falcon III

Number of Movies

2895

File Format

TIFF-LZW

Frame Size

4096 × 4096

Frames per Movie

64

Pixel Size

0.835Å

Particle Extraction Box Size

420 × 420

Benchmarks

Metric

Value

Movies Pre-processed Per Hour Per GPU

493

Movies Pre-processed Per Day Per GPU

11832

Average Pre-processing Time Per Movie

7.3s

Falcon IV EER (Apoferritin)

Benchmark results for ~3000 Electron Event Representation (EER) movies from a Falcon IV detector. The particle is highly symmetric. The target apoferritin is highly symmetric. Enough information is present in the dataset to approach atomic resolution.

Particles were selected with the ring template picker strategy. Streaming 2D Classification, Ab-initio Reconstruction and Streaming Refinement yielded a 1.9Å resolution map from ~700,000 particles without any additional processing.

Data Properties

Property

Value

Detector

TFS Falcon IV

Number of Movies

3370

File Format

EER

Frame Size

8192 × 8192

Frames per Movie

434 (40 used)

Pixel Size

0.457Å

Particle Extraction Box Size

512 × 512

Benchmarks

Metric

Value

Movies Pre-processed Per Hour Per GPU

303

Movies Pre-processed Per Day Per GPU

7272

Average Pre-processing Time Per Movie

11.9s

K2 TIFF with Preprocessing + 2D/3D Streaming (CB1 Gi)

Benchmark results for ~3000 TIFF-LZW compressed movies from a GATAN K2 detector. The target complex is a small, flexible membrane protein.

Particles were selected with the Template Picker strategy. Streaming 2D Classification, Ab-initio Reconstruction and Streaming Refinement yielded a 3.9Å resolution map from ~700,000 particles.

Data Properties

Property

Value

Detector

Gatan K2

Number of Movies

2756

File Format

TIFF-LZW

Frame Size

3838 × 3710

Frames per Movie

40

Pixel Size

0.86Å

Particle Extraction Box Size

360 × 360

Particle Extraction Bin Size

256 × 256

Benchmarks

Pre-processing and streaming results for this dataset measured with Hardware Configuration 3

Metric

Value

Movies Pre-processed Per Hour Per GPU

870

Movies Pre-processed Per Day Per GPU

20880

Average Pre-processing Time Per Movie

4.13s

Post-Processing: 256 × 256 box size, 50 2D Classes

Metric

Value

Particles Extracted for 2D Classification

79,278

Time to 2D Classify Extracted Particles

4 minutes, 16 seconds

Particles Used for Reconstruction

100,000

Time to Reconstruct Initial Volume (Ab-initio)

5 minutes, 11 seconds

Particles Selected for Refinement

278,312

Time to Refine Final Volume

22 minutes, 55 seconds

Last updated