With RAM price inflation, I wonder if FPGAs can be used to offload inference processing without keeping weights in RAM? The available RAM would be for activations, KV Cache, context but not static weights. Weights could be streamed from disk. This approach is not for tokens/second but throughput at a lower cost. Possibly better answers/kHh? I've started researching this, but wonder if others have considered/tried this?