We use the chip general purpose. HFT simulation code, empirically faster. The GPUs are to train a model. Separate use-case, but we use same hardware for both u

We use the chip general purpose. HFT simulation code, empirically faster. The GPUs are to train a model. Separate use-case, but we use same hardware for both use-cases so that's why the big cache thing.