Compute requirements

Atacformer is a transformer model that leverages a large context-window, due to the nature of scATAC-seq data. A typical Atacformer context-window size is 8,192 tokens. To that end, Atacformer requires a significant amount of GPU memory to use in its full capacity.

Atacformer processes data in chunks (batches). This batch_size is configurable and can be adjusted based on the available hardware resources and the specific requirements of the task at hand. You can use the following table as a general rule of thumb for how much GPU VRAM you need:

Batch Size	GPU VRAM Required	Example GPUs
4	2GB	NVIDIA RTX 2080 Ti
8	4GB	NVIDIA RTX 3080
16	8GB	NVIDIA RTX 3090
32	16GB	NVIDIA RTX 4070 Ti
64	48GB	NVIDIA A40/A6000