Most legacy NPU accelerators were highly optimized for the ResNet structures that were state-of-the-art five years ago. Transformers for vision and LLM applications are vastly different in network topology and building block operators. If your legacy NPU cannot handle Transformer networks it's time to switch to Quadric Chimera GPNPUs. Chimera GPNPUs can run any network and support any ML graph operator - even today's newest Transformers.
Try Devstudio now
Sign-in to your existing account, or sign-up for a new Quadric DevStudio account today and see it for yourself.