dorado
dorado copied to clipboard
nanoporetech
Request: Document GPU Performance in a Table
To determine which GPUs people need for their projects it would be great to have some benchmarks for the GPUs that you test dorado on. If you expand the table you already have with some values for the amount of data that can be basecalled/hour or something that would be really helpful.
| Platform | GPU/CPU | Minimum Software Requirements | fast (Gbp/hour) | hac (Gbp/hour) | sup (Gbp/hour) | sup all mods (Gbp/hour) |
|---|---|---|---|---|---|---|
| Linux x86_64 | (G)V100 | CUDA Driver ≥450.80.02 | ||||
| A100 | CUDA Driver ≥450.80.02 | |||||
| H100 | CUDA Driver ≥520 | |||||
| Linux arm64 | Jetson Orin | Linux for Tegra ≥34.1.1 | ||||
| Windows x86_64 | (G)V100 | CUDA Driver ≥452.39 | ||||
| A100 | CUDA Driver ≥452.39 | |||||
| H100 | CUDA Driver ≥520 | |||||
| Apple | Apple Silicon (M1) | |||||
| Apple Silicon (M2) |
The reference to the old benchmarks indicate that it is somewhat outdated (dorado 0.3.0)) and no H100. https://aws.amazon.com/blogs/hpc/benchmarking-the-oxford-nanopore-technologies-basecallers-on-aws/
I would add that it would be incredibly helpful to also include top-end consumer cards, like the RTX 4090. Often these are the most cost effective based on their performance.
Perhaps if you make a benchmarking protocol, people can run it and post the results. I'd be happy to run it on all of our machines.
James
How about we derive a benchmark using that data? https://nanoporetech.com/document/telomere-to-telomere-sequencing-t2t-on-promethion-sqk-apk114-sqk#downstream-analysis
