I first saw Tim’s post when the Ampere series was released in 2020. It gave practical recommendations based on the workflow type and scale of deep learning research. Many things have changed since then: transformer architectures and large language models (LLMs) became mainstream, a new generation of GPUs was launched with FP8 support, the consumer/prosumer card shortage situation has improved… The latest update incorporated detailed analysis and suggestions based on these new developments. Tim is also kind enough to include a GPU recommendation chart in the updated post, and that is the source of the decision tree data for this tool.
Note that the tool and the original chart do not cover every situation. It is likely that one still needs to evaluate specific hardware requirements on a case-by-case basis. For example, running inference on protein sequences using the AlphaFold model (which also uses a transformer architecture) is heavily memory-bound. To me, the primary considerations would be
- the protein sequence lengths and the required GPU memory size,
- the model and implementation used, for example, alternatives such as AlphaFold-Multimer or OpenFold, and
- the fact if we want to run such workflows on-demand or routinely.
That’s it! I hope the tool is useful for choosing your next card.