> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/clara/parabricks/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/clara/parabricks/_mcp/server.

# 4.5.0-1 Release Notes

**Highlights:**

* Support for the following NVIDIA Blackwell GPU architectures: SM\_100 and SM\_120.
* As of v4.5 we are dropping support for the Volta (V100) line of GPUs.
* Splice-aware support and performance improvements for [minimap2](/tool-reference/tools/minimap-2).
* Performance improvements for [giraffe](/tool-reference/tools/giraffe-vg-giraffe-gatk).
* Marking duplicates is now on by default for [giraffe](/tool-reference/tools/giraffe-vg-giraffe-gatk).
  Use `--no-markdups` to turn off marking duplicates.
* Faster [rna\_fq2bam](/tool-reference/tools/rna-fq-2-bam) with `--num-streams-per-gpu`.
* Performance improvements for force-calling mode in [haplotypecaller](/tool-reference/tools/haplotypecaller) and
  [mutectcaller](/tool-reference/tools/mutectcaller).

## Improvements

### Tool Updates

[fq2bam](/tool-reference/tools/fq-2-bam-bwa-mem-gatk) and [fq2bam\_meth](/tool-reference/tools/fq-2-bam-meth):

* Added `--in-fq-list` and `--in-se-fq-list` support. You can specify
  a file containing a list of paired-end or single-end FASTQ file paths along
  with their corresponding read group information.
* Recovery mode has been improved to avoid falling back to CPU—enhancing performance. This may
  happen often when an input FASTQ produces a large amount of SMEMs (SuperMaximal Exact Matches) per read.
* Peak host memory use can be indirectly modified by changing `--bwa-normalized-queue-capacity`.
* Improved performance on devices with compute capability
  SM\_90 (e.g. H100, H200) and SM\_100 (e.g. B200).

[giraffe](/tool-reference/tools/giraffe-vg-giraffe-gatk):

* Improved performance across the board.
* Marking duplicates is now enabled by default. The option `--no-markdups` has
  been added to skip the marking duplicates step.
* Added support for `--ref-paths`. This option allows to specify an ordered
  list of paths in the graph and helps supporting pangenome graphs
  containing multiple reference paths in variant calling pipelines.
* Added `--in-fq-list` and `--in-se-fq-list` support. You can specify
  a file containing a list of paired-end or single-end FASTQ file paths along
  with their corresponding read group information.
* Added `--copy-comment` which allows the user to copy FASTQ comments to
  BAM output via the auxiliary tag. Similar to `-C` in
  [fq2bam](/tool-reference/tools/fq-2-bam-bwa-mem-gatk) and BWA-MEM.

[markdup](/tool-reference/tools/markdup):

* Added `--markdups-single-ended-start-end` to mark duplicate on single-ended reads
  by 5' and 3' end.

[minimap2](/tool-reference/tools/minimap-2):

* Added support for `--preset` options `splice` and `splice:hq`.
* Performance improvements.
* Added `--max-queue-reads` and `--nstreams` performance options.

[rna\_fq2bam](/tool-reference/tools/rna-fq-2-bam):

* Added `--num-streams-per-gpu` to support multiple GPU streams.

[mutectcaller](/tool-reference/tools/mutectcaller):

* Faster force-variant call `--mutect-alleles` in mutectcaller.

[haplotypecaller](/tool-reference/tools/haplotypecaller):

* Faster force-variant call `--htvc-alleles` in haplotypecaller.

[deepvariant](/tool-reference/tools/deepvariant):

* Added support for `--keep_legacy_allele_counter_behavior`.

[deepvariant](/tool-reference/tools/deepvariant):

* Added `--use-tf32` to utilize tensor cores during inference to achieve better performance on
  Ampere+ GPUs. Note the accuracy might be slightly impacted because of tf32 lower precision.

### General

* Support for the following NVIDIA Blackwell GPU architectures: SM\_100 and SM\_120.
* Improved error messaging when Parabricks receives signals from the OS, such as out-of-memory (OOM) killer events.

## Bug Fixes

[fq2bam\_meth](/tool-reference/tools/fq-2-bam-meth):

* Fixed crash which occurred when the input FASTQ did not contain read comments.

[haplotypecaller](/tool-reference/tools/haplotypecaller):

* Fixed a hash table overflow in De Bruijn graph.

[markdup](/tool-reference/tools/markdup):

* Fixed a crash which occurred when the input BAM did not have a read group line. Error will be handled at the start.
* Fixed correctness issue when the read queryname had non-numeric characters in the colon-delimited end of the name.

For further information refer to the
[Parabricks datasheet](https://resources.nvidia.com/en-us-hc-genomics/healthcare-genomics).