We ran Artemis against OpenAI's Whisper on Intel's AI Tiber Cloud, targeting Xeon 8380P processors with Gaudi accelerators.
The setup
The constraint was deliberately strict: the model architecture does not change. No swapping layers, no quantisation trade-offs presented as free wins, no retraining. Only the code around the model is allowed to move.
No architecture changes
Under that constraint Artemis found 20% faster inference. The value of the constraint is that the result is boring to adopt — there is no accuracy conversation to have with a risk committee, because the model is byte-for-byte the one you already approved.
