Back to Blog
![]() ![]() Direct Hardware Mapping of Dot Products in convolution layers: A Binary Adder tree sums the partial-products 3. Number of MOAs and number of mean non-null inputs per Adder in AlexNet layers However, for state-of-the-art CNNs, C,J,K can be large, leading to adders with up to 1774 operands (cf table 1.) As a result, most of the logic required to map a CNN layer is dedicated to the MOAs part, which corresponds, for instance, to 69% of the resources in the first layer of an AlexNet. By default, synthesis tools instantiate deep binary adder trees 2 2 2Binary adders refer to adders with TWO operands and NOT adders with a 1-bit operand to implement MOAs, which require C J K − 1 binary adders. A drawback of the DHM solution is that each layer requires N Multiple Operand Adders with C × J × K inputs in order to accumulate the partial products.
0 Comments
Read More
Leave a Reply. |