I've read elsewhere that it's 10x less electricity for inference workloads compared to standard GPUs. It is not clear to me, are the model weights built into the silicon (e.g. per model tapeout), or is this a new kind of chip architecture that still has weights in DRAM/SRAM?