Enabling fully connected probabilistic computing through a fast pipelined multi-operand adder

Probabilistic computing has gained a prominent role as combinatorial optimization solvers on classical hardware. Probabilistic bits (pbits) must be updated sequentially to rapidly converge to the lowest energy state of the objective function. Yet, all the current implementations rely on mappings to sparse graphs to speed up the update operation. In this paper, we present a new pipelining technique for a multi-operand adder to enable a fully connected structure of the pbit native graph. Previous approaches lack a pipelined architecture and avoid data dependencies by updating each pbit only after all its connected pbits are updated. This results in a small adder but limits execution to sparse problems and implies a prominent supplement of pbits due to sparsification. Our implementation uses a fast pipelined adder that receives new operands at each pipe stage, handling data dependencies during execution. The pipelined unit was evaluated against other pipeline strategies and state-of-the-art emulator implementations, demonstrating promising pbits update times. The obtained performance closely matches implementations on sparse graphs, implying greater scalability, especially for denser problems. With a future parallelized architecture, this may enable fast probabilistic computing for fully connected problems, avoiding a significant or even unacceptable increase in the number of required pbits