Validated observations

Performance observations with clearly defined limits

The values on this page come from SHA-256-verified analysis records. They were obtained on one validated workstation using fixed input data and should not be interpreted as a general CPU/GPU performance comparison.

GPU alignment of the complete ERR188044 dataset

For the complete ERR188044 dataset, observed peak host memory use was 49.35 GiB with one-pass alignment and 57.42 GiB with two-pass alignment. — The public ERR188044 input contained 36,349,964 paired fragments and was analyzed on Ubuntu 24.04.4 with an RTX 3090 (24 GiB VRAM) and 128 GB RAM. One completed analysis was checked for each profile. The isolated two-pass Parabricks step took 850 seconds. An isolated one-pass time is unavailable because the final verification reused the cached alignment.

Profile	Computational resources	Observed peak host memory	Observed peak VRAM	Isolated alignment time	Result
One-pass	42 GB / 12 CPU cores	52,991,762,432 B (49.35 GiB)	11,598 MiB	Not measured separately	Completed successfully
Two-pass Basic	96 GB / 12 CPU cores	61,657,542,656 B (57.42 GiB)	11,606 MiB	850 s (14 min 10 s)	Completed successfully; cgroup peak 80,275,611,648 B

These measurements were obtained on one validated workstation using one fixed public dataset. They are not a matched CPU STAR benchmark because Harako-RNAseq runs a different workflow, and they should not be interpreted as a general GPU speedup estimate.

Salmon runtime

The median runtime was 28.200 seconds for Salmon 1.10.3 and 22.658 seconds for Salmon 2.5.1. — The ERR188044 C1 1M input contained 958,978 processed paired fragments. Both versions ran on the same workstation with source-compatible version-specific indices and six threads. Each version was run twice, with the execution order reversed. These n=2 results are descriptive and do not establish a general performance difference.

Version	Median wall time	Repeatability
1.10.3	28.1998923805 s	Repeated runs were highly concordant, but the numerical results were not identical.
2.5.1	22.6584244570 s	Repeated runs produced identical numerical results.

The observed median wall-time ratio for 1.10.3 relative to 2.5.1 was 1.2445654566. Because n=2, this should not be interpreted as a general performance estimate.

Agreement in abundance estimates between Salmon versions

Median Spearman correlations between Salmon versions were 0.953697 for transcript TPM, 0.987481 for gene TPM, 0.954389 for transcript NumReads, and 0.988910 for gene NumReads. — These two order-reversed comparison runs used the same C1 input and workstation. The table reports Q1, Q2, and median Spearman correlations. High correlation does not show that results from the two versions are interchangeable or that either agrees with biological ground truth.

Measure	Comparison run Q1	Comparison run Q2	Median
Transcript TPM	0.9537426073	0.9536519157	0.9536972615
Gene TPM	0.9874856268	0.9874762280	0.9874809274
Transcript estimated reads (NumReads)	0.9544253338	0.9543532364	0.9543892851
Gene estimated reads (NumReads)	0.9889113990	0.9889087428	0.9889100709

Genes and transcripts affected by the Salmon version difference

Counts above the predefined difference threshold in comparison runs Q1 and Q2 were 2,882 and 2,883 for transcript TPM, 733 and 731 for gene TPM, 2,929 and 2,925 for transcript NumReads, and 552 and 554 for gene NumReads. — For the C1 input, this figure shows the number of genes or transcripts that exceeded a predefined descriptive difference threshold in two order-reversed comparisons. These counts may reflect differences in version, implementation, index construction, or execution conditions. They do not show that either version is incorrect or differs from biological ground truth.

Measure	Comparison run Q1	Comparison run Q2
Transcript TPM	2,882	2,883
Gene TPM	733	731
Transcript estimated reads (NumReads)	2,929	2,925
Gene estimated reads (NumReads)	552	554

Salmon 2.5.1 is the default for new analyses. Salmon 1.10.3 remains available for compatibility with earlier analyses. Repeated 1.10.3 runs were highly concordant, but their numerical results were not identical. Results produced with different Salmon versions should not be combined without explicitly accounting for the version difference. Every run records the version, container image, index, options, and execution order.