1. NP-complete problems are (today) global search problems (3-SAT algos search for a satisfying assignment...);
2. Combinatorial search spaces are not only non-convex, they don't have gradients (sub-gradients don't count) and even if they did, without convexity, it's still hopeless.
Thus, you cannot (today) appreciably parallelize NP-complete algos, you can only draw more lottery tickets (for where to seed your search), i.e., constant factor improvement.
I'm just putting this out there in case someone gets tantalized by the "fancy math/physics/fpga thing solves NP-complete problems" hook. Yes it can but no better than your computer (despite what paper authors claim in order to get published). To wit: if this were a compelling solution to NP-complete problems it would either a) already be a proof that NP=P or b) be implemented in absolutely every single SAT solver (using GPU or whatever rather than FPGA). In contrast, every single SAT solver uses CDCL which is ~roughly DFS with clever backtracking.
And I say this as someone that has investigated solving ILPs on FPGA - i.e., I wish it were true but it ain't.