Performance

How PowerFit performs on different computational resources, such as CPU and GPU, and with different batch sizes. This can help users understand the trade-offs between different configurations and choose the best one for their needs.

Measurements

Fetch the map and the structure for the test case:

wget https://ftp.ebi.ac.uk/pub/databases/emdb/structures/EMD-1046/map/emd_1046.map.gz
wget https://files.rcsb.org/download/9A2G.cif.gz

Runs on machines with the following specifications:

• m1: AMD Ryzen 5 5600G and NVIDIA GeForce RTX 3050
• m2: AMD Ryzen 7 7800X3D and AMD Radeon RX 7900 XTX
• m3: AMD EPYC 9554 and NVIDIA RTX 6000 Ada
• m4: Intel i7-13700H and NVIDIA RTX 4050 Laptop via WSL

Run commands to test different computational resources

# On machine 1
mkdir -p runs
for run in 1 2 3 4 5; do
    for batch_size in 1000 500 250 200 150 100 75 50 10 0; do
        powerfit emd_1046.map.gz 20 9A2G.cif.gz -a 4.71 --delimiter , -n 0 --gpu cuda:0 --batch-size $batch_size -d runs/m1-cuda-bs${batch_size}-r${run}
    done
    for batch_size in 1000 500 250 200 150 100 75 50 10 0; do
        powerfit emd_1046.map.gz 20 9A2G.cif.gz -a 4.71 --delimiter , -n 0 --gpu 0:0 --batch-size $batch_size -d runs/m1-opencl-bs${batch_size}-r${run}
    done
done
# On machine 2
mkdir -p runs
for run in 1 2 3 4 5; do
    powerfit emd_1046.map.gz 20 9A2G.cif.gz -a 4.71 --delimiter , -n 0 --nproc 6 -d runs/m2-cpu6-r${run}
    powerfit emd_1046.map.gz 20 9A2G.cif.gz -a 4.71 --delimiter , -n 0 --nproc 6 --progressbar -d runs/m2-cpu6-pb-r${run}
    for batch_size in 3500 1000 250 200 150 100 75 50 10 0; do
        powerfit emd_1046.map.gz 20 9A2G.cif.gz -a 4.71 --delimiter , -n 0 --gpu 0:0 --batch-size $batch_size -d runs/m2-opencl-bs${batch_size}-r${run}
    done
done

# On machine 3
mkdir -p runs
for run in 1 2 3 4 5; do
    for batch_size in 1000 500 250 200 150 100 75 50 10 0; do
        powerfit emd_1046.map.gz 20 9A2G.cif.gz -a 4.71 --delimiter , -n 0 --gpu cuda:0 --batch-size $batch_size -d runs/m3-cuda-bs${batch_size}-r${run}
    done
    for batch_size in 4000 3000 2000 1000 500 250 200 150 100 75 50 10 0; do
        powerfit emd_1046.map.gz 20 9A2G.cif.gz -a 4.71 --delimiter , -n 0 --gpu 0:0 --batch-size $batch_size -d runs/m3-opencl-bs${batch_size}-r${run}
    done
done

# On machine 4
mkdir runs
for run in 1 2 3 4 5; do
    for batch_size in 1000 500 250 200 150 100 75 50 10 0; do
        powerfit emd_1046.map.gz 20 9A2G.cif.gz -a 4.71 --delimiter , -n 0 --gpu cuda:0 --batch-size $batch_size -d runs/m4-cuda-bs${batch_size}-r${run}
    done
done

Convert logs to table

python3 docs/parse_times.py "runs/*/*.log" > docs/times.csv
python3 docs/batch_size_plot.py docs/times.csv docs/batchsize_vs_search.csv

Group by and markdown table with duckdb

.mode markdown
WITH runs AS (
  SELECT
    regexp_replace(run, '-r[0-9]+$', '') AS run_group,
    total_seconds::DOUBLE AS total,
    search_seconds::DOUBLE AS search,
    batch_size
  FROM read_csv_auto('docs/times.csv')
)
SELECT
  run_group,
  round(avg(search), 2) AS avg_search,
  round(stddev_samp(search), 2) AS stddev_search,
  min(search) AS min_search,
  max(search) AS max_search,
  median(search) AS median_search,

  round(avg(total), 2) AS avg_total,
  round(stddev_samp(total), 2) AS stddev_total,
  min(total) AS min_total,
  max(total) AS max_total,
  median(total) AS median_total,

  min(batch_size) AS batch_size,
  count(*) AS nr_runs
FROM runs
GROUP BY run_group
ORDER BY run_group;

The times.csv contains the parsed measurements taken around 30 April 2026 on commit 0e60abd4f69d3d438ddaee0651519a79d99fa0f3 of code. The batchsize_vs_search.csv file is a chart-ready view with normalized run labels and preserved run_number values for replicate spread.

Batch size impact

{ "$schema": "https://vega.github.io/schema/vega-lite/v6.json", "title": "Batch Size vs Search Seconds", "data": { "url": "../batchsize_vs_search.csv", "format": { "type": "csv" } }, "transform": [ { "filter": { "field": "batch_size", "lt": 1000 } } ], "params": [ { "name": "series", "select": { "type": "point", "fields": ["run"] }, "bind": "legend" }, { "name": "grid", "select":"interval", "bind":"scales" } ], "mark": {"type": "point", "tooltip": true}, "encoding": { "x": { "field": "batch_size", "type": "quantitative", "title": "Batch Size" }, "y": { "field": "search_seconds", "type": "quantitative", "title": "Search (s)" }, "color": { "field": "run", "type": "nominal", "title": "Run", "scale": { "scheme": "category10" } }, "opacity": { "condition": { "param": "series", "value": 1 }, "value": 0.01 }, "tooltip": [ { "field": "run", "type": "nominal", "title": "Run" }, { "field": "run_number", "type": "ordinal", "title": "Run Number" }, { "field": "batch_size", "type": "quantitative", "title": "Batch Size" }, { "field": "search_seconds", "type": "quantitative", "title": "Search (s)" } ] } }

Based on plots, the default batch size is set to 100.

When batch size is set 0 then the rotations are processed one by one instead of in batches.

When you have very few rotations (--angle 20) setting batch size to 0 can be faster.

Legend

• cpuN: CPU with N processes
• pb: Run with progress bar enabled
• opencl: GPU with OpenCL backend
• cuda: GPU with CUDA backend
• autobs: GPU with automatic batch size and the respective backend
• nobs: GPU with serial rotations and the respective backend
• bsNNNN: GPU with batch size of NNNN and the respective backend
• rN: Run number N
• total: Total time taken for the run, includes reading input, writing output, and all computations
• search: Time taken for the all computations
• batch_size: Batch size used for the run