The Save Argmax option enables argmax and max feature along channel dimension when restoring the outputs back to DDR space. In some cases like segmentation, only the index of the maximum value is needed. Then it is useful to replace softmax with argmax in the model to remove Exp calculation and reduce latency.
DPUCZDX8G Architecture | Extra LUTs | Extra Registers |
---|---|---|
B512 | 422 | 556 |
B800 | 399 | 547 |
B1024 | 460 | 546 |
B1152 | 503 | 631 |
B1600 | 590 | 640 |
B2304 | 803 | 442 |
B3136 | 832 | 758 |
B4096 | 735 | 389 |