A Reconfigurable lsing Machine for Boolean Satisfiability Problems Featuring Many-Body Spin Interactions
Ising machines [1-6] have recently shown their unique capabilities in solving hard combinatorial optimization problems (COPs) intractable using conventional computers. We map the COPs to the Ising model implemented in a network of hardware spins (i.e., computing units of the Ising machine) and find...
Uloženo v:
| Vydáno v: | Proceedings of the Custom Integrated Circuits Conference s. 1 - 2 |
|---|---|
| Hlavní autoři: | , , |
| Médium: | Konferenční příspěvek |
| Jazyk: | angličtina |
| Vydáno: |
IEEE
01.04.2023
|
| Témata: | |
| ISSN: | 2152-3630 |
| On-line přístup: | Získat plný text |
| Tagy: |
Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
|
| Shrnutí: | Ising machines [1-6] have recently shown their unique capabilities in solving hard combinatorial optimization problems (COPs) intractable using conventional computers. We map the COPs to the Ising model implemented in a network of hardware spins (i.e., computing units of the Ising machine) and find optimal solutions by observing the natural convergence of binary states of spins based on interactions between them. The existing Ising machines [3-5] have implemented hardware spins that only interact with their neighbors, lacking the generality to solve more complicated COPs [1] with arbitrary interactions between spins. A prior work [6] proposed a flexible circuit to map an arbitrary interaction between remote spins. While it resolves the overhead of mapping complicated Ising models to the hardware, the mapping has been limited to the Ising model with two-body spin interactions, while there are many COPs, such as Boolean satisfiability (SAT), requiring N-body interactions (where N > 2). Approximate algorithms have been used to conveit the problem with many-body spin interactions to the Ising model with two-body spin interactions. However, they result in a significant hardware overhead with many ancillary spins and lower accuracy due to the approximation. Recent works [1-2] implemented many-body interactions in quantum and optical machines. However, they suffer from limited scalability and require high development and operation cost. in this paper, we present a CMOS Ising machine with reconfigurable processing elements to solve complex combinatorial optimization problems mapped to the model with more than two-body spin interactions. The popular 3SAT problems and their variants are demonstrated using 128 reconfigurable processing elements (PEs) integrated into a 65nm prototype chip. Each PE consists of eight spins with up to 56 temporally reconfigurable spin interactions based on a King's graph topology. The proposed Ising machine computes massively parallel spin operations in a complex spin network based on multi-cycle spin operations. |
|---|---|
| ISSN: | 2152-3630 |
| DOI: | 10.1109/CICC57935.2023.10121303 |