Designed for multiple, simultaneous users, DGX Station A100 leverages server-grade components in an office-friendly form factor. It has four fully interconnected and Multi-Instance GPU (MIG)-capable NVIDIA A100 Tensor Core GPUs with 320 GB of total GPU memory that can plug into a standard power outlet.

• AI workgroup server delivering 2.5 petaFLOPS of performance that your team can use without limits — for training, inference, and data analytics
• Server-grade, plug-and-go, without requiring data center power and cooling
• World-class AI platform, with no complicated installation or IT help needed
• The world’s only workstation-style system with four fully interconnected NVIDIA A100 Tensor Core GPUs and 320 gigabytes (GB) of GPU memory
• Delivers a fast-track to AI transformation with NVIDIA know-how and experience

The DGX A100 station provides users with a simple solution to expand GPU resource to run GPU-enabled tools. Provided with installation instructions and optimal running parameters for running production basecalling tools, users can expand their post-run basecalling capacity. Installation instructions and optimal running conditions for other GPU-enabled Oxford Nanopore tools will be made available. Users can run other tools at their discretion with support from NVIDIA.

For more information, please see the NVIDIA DGX Station A100 datasheet.

Alternatively, you can find installation instructions online at https://docs.nvidia.com/dgx/dgx-station-a100-qsg/index.html#abstract

After purchase of the device and 3-year enterprise support package, you should receive an email from NVIDIA to the account that performed the purchase with details on how to setup your support account with NVIDIA.

If you have any problems with the DGX Station A100, please contact NVIDIA enterprise support for assistance: https://www.nvidia.com/en-us/support/enterprise/ . We suggest you copy your local Field Application Specialist into your communication with NVIDIA enterprise support.

To use Guppy on the DGX Station A100, install the GPU version of the software on the station from the Debian package as described in the Guppy protocol.

1. Use the Guppy basecall server, so that there is a single program managing Guppy's GPU use.
2. Use multiple processes to overcome issues which arise when a single process is unable to read .fast5 files fast enough.

1. Basecalling a single folder with a lot of data in it.
2. Basecalling many data folders in separate tasks, as if a job scheduler is being used.

1. Starting a new basecall server.
2. Using the basecall server for basecalling, in one of the two use cases outlined above.
3. Shutting down the server when basecalling is complete.

Guppy ships with a "basecaller supervisor" that will launch many basecalling clients in parallel, to mitigate issues with reading .fast5 files. Each client has an ID, numbered upwards from zero, which allows their output files to be created in parallel.

If using the Fast basecall model, we recommend to set num_clients to 50. For all other models, set num_clients to 20.

--- /save_folder/
    | fastq_runid_6dce0a5_client0_0_0.fastq
    | fastq_runid_6dce0a5_client1_0_0.fastq
    | fastq_runid_6dce0a5_client2_0_0.fastq
    | guppy_basecaller_0_log-2019-11-25_15-11-53.log
    | guppy_basecaller_1_log-2019-11-25_15-11-53.log
    | guppy_basecaller_2_log-2019-11-25_15-11-53.log
    | guppy_basecaller_supervisor_log-2019-11-25_15-11-53.log
    | sequencing_summary_0.txt
    | sequencing_summary_1.txt
    | sequencing_summary_2.txt
    | sequencing_telemetry_0.js
    | sequencing_telemetry_1.js
    | sequencing_telemetry_2.js

The most scalable way to basecall many data folders is to use a single basecall client for each data folder.

Below are approximate basecall speeds, in Gbases per hour, that you should be able to attain using the Guppy setup outlined above. Actual speeds will vary depending on the type of data you have: for example, shorter reads will basecall more slowly as they are less efficient to move through the basecall server and process on the GPU. Basecall speeds will also decrease if operations are requested that require additional processing time, such as barcoding or alignment.

The NVIDIA DGX Station A100 has 7.68 TB of internal data storage. Data flow should be managed accordingly to best suit the needs of the end users. The location and volume of the data, network connection speed as well as the basecalling model selected will all have an impact on optimal data flow.

100 Gbases of sequencing data in .fast5 format typically occupies approximately 1 Tbyte of storage. Variables such as read length will alter this ratio.

For the fastest basecalling speed, .fast5 data should be stored locally on the DGX Station A100. Depending on the basecalling model used and the networked storage available, users could consider basecalling data from networked storage. The table below compares the % basecalling speed vs model speed for basecalling data from local SSD storage when Guppy is run with the suggested parameters.

Note that these benchmarks are for the previous model of the DGX Station A100 (160 GB). Benchmarks for the 320 GB model will be released soon.

If you choose to basecall from .fast5 files locally on the DGX Station A100 to reduce time copying .fast5 data, we recommend creating only a temporary copy of the .fast5 data on the DGX Station A100 and deleting the .fast5 data after basecalling. Move only the analysed (FASTQ) data off the station.