Abstract
The model used for architectural optimization is discussed. It is then shown that the optimal architecture for 2-D convolution uses a 1-D convolution combinatorial operator to reduce the 2-D problem to a classical 1-D systolic array. Furthermore, for a given technology and chip area, if one tries to apply the proposed cost function to this architecture, it is possible to predict a practical bound for the kernel size of convolutions that can be implemented on one chip. Details are given of a 5 × 10 real-time convolver chip that exhibits several new features, including a multikernel mode.
| Original language | English |
|---|---|
| Journal | Proceedings of the Custom Integrated Circuits Conference |
| Publication status | Published - 1 Dec 1990 |
| Event | Proceedings of the 12th Annual IEEE 1990 Custom Integrated Circuits Conference - CICC '90 - Boston, MA, USA Duration: 13 May 1990 → 16 May 1990 |