NVIDIA Docs Hub Homepage NVIDIA Clara Welcome to NVIDIA Parabricks v4.7.1 Improvements

Improvements

All tools that rely on Parabricks GPU-accelerated sort now have improved error handling. Running pbrun <tool> --help or pbrun <tool> -h now works without GPUs visible.

Tool Updates

deepvariant, deepvariant_germline, pacbio_germline, and ont_germline:

Improved performance with the small model across WGS, PacBio, and ONT modes.
The small model is now enabled by default. Use --disable-small-model to disable it.

pangenome_aware_deepvariant:

Added Roche SBX mode with --sbx for GPU-accelerated variant calling on Roche SBX-D and SBX-Fast short-read data.

giraffe:

Improved performance of the GPU seed-extension (GBWT) stage.
Improved performance when producing final read alignments.
Improved performance of the GPU Dozeu alignment kernel.
Improved initialization time for single-end alignment.
Moved minimizer payload decoding to the GPU, reducing CPU load.
Additional performance improvements and reduced host-side memory usage during alignment.

minimap2:

Updated version of minimap2 to v2.31.
Added --chaining-on-gpu flag to use GPUs to accelerate chaining, a compute-intensive step in minimap2 alignment, reducing CPU load.
Improved performance of GPU-accelerated spliced alignment.

fq2bam and fq2bam_meth:

Improved performance for both single-end and paired-end alignment.
Added partition mode (--run-partition), which runs one process per partition to more efficiently split work across multiple GPUs. The --bwa-gpu-num-per-partition option controls the number of GPUs each partition drives (default: 2).
Increased the maximum read size supported for processing on GPU to 1024 bp. The maximum read size to process on GPU can be controlled via --max-read-size-gpu, where valid values are 512 or 1024 bp and the default remains 512 bp. In either mode, larger reads will continue to go to CPU recovery.

fq2bam_meth:

Updated the default CPU threading configuration to values that are more optimal for end-to-end runtime: --bwa-primary-cpus now defaults to 16 * <number of GPUs> and --bwa-cpu-thread-pool now defaults to 1. Both options can still be set manually to override the defaults.

rna_fq2bam:

Performance improvement through code optimization.
Added throughput metrics.

haplotypecaller and mutectcaller:

Added multi-stream support with the --num-streams-per-gpu option, which controls the number of streams to use per GPU (default: 1).

bamsort, deepsomatic, deepvariant, haplotypecaller, minimap2, mutectcaller, and pangenome_aware_deepvariant:

Added support for reading CRAM v3.1 file.