Table 3.
Genomic relationship matrix calculation times (sec)
| Dataset | Machine | GCTA | PLINK 1.90 | Ratio |
|---|---|---|---|---|
| synth1p | Mac-2 | 222.2 | 7.2 | 31 |
| Mac-12 | 184.7 | 5.0 | 37 | |
| Linux32-8 | 248.4 | 10.9 | 22.8 | |
| Linux64-512 | 4.4 | 9.6 | 0.46 | |
| Win32-2 | 373.1 | 39.3 | 9.5 | |
| Win64-2 | 367.2 | 6.6 | 56 | |
| synth2p | Mac-2 | nomem | 805.8 | |
| Mac-12 | 17.0 k | 138.3 | 123 | |
| Linux32-8 | nomem | 1153.4 | ||
| Linux64-512 | 65.1 | 318.9 | 0.20 | |
| Win32-2 | nomem | 2007.2 | ||
| Win64-2 | nomem | 450.1 | ||
| chr1snp | Mac-2 | nomem | 87.1 | |
| Mac-12 | 2260.9 | 50.9 | 44.4 | |
| Linux32-8 | nomem | 94.3 | ||
| Linux64-512 | 58.3 | 91.6 | 0.64 | |
| Win32-2 | nomem | 317.5 | ||
| Win64-2 | nomem | 65.7 |
This involves a variance-normalizing distance function which cannot be efficiently computed with just bit population counts. PLINK 1.9’s lookup table-based algorithm is slower than GCTA 1.24 on heavily multicore machines (see the Linux64-512 results), but has complementary advantages in portability, accuracy, and memory efficiency.