Changing the Game: The Bounce-Bind Ising Machine

Abstract

The Ising model, originally proposed a century ago, has become a cornerstone of combinatorial optimization in recent decades. However, Ising machines remain constrained by a fundamental hardware-speed trade-off. We introduce the Bounce-Bind Ising Machine (BBIM), a mechanism with a single tunable parameter that modulates spin dynamics without altering the energy landscape, building upon the classic golf-ball analogy but replacing it with a dynamic tennis ball/shot put system. The Bounce mode (accelerating escapes from local minima) and Bind mode (enabling rapid convergence) dynamically balance speed and quality. Benchmarked on dense MAX-CUT (edge density=0.5), BBIM achieves a peak speedup of 6.15 times at n=200. For sparse 3-Regular 3-XORSAT (second-order), the peak speedup reaches 27.3 times at n=160. Both results incur negligible additional hardware resource consumption. This work demonstrates a critical pathway to circumventing the hardware-speed bottleneck and its practical applicability to large-scale optimization hardware, validated on structurally distinct benchmarks.

Figures (8)

• Figure 1: State transition matrices under different Bounce-Bind parameters in a case of a 3-bit Ising problem. The results show that the probability of the system transitioning to its own state increases monotonically with the parameter $\mathcal{B}$, which influences the convergence of the system. The state transition probability is calculated by the algorithm in Supplementary Section S1.
• Figure 2: Probability distribution function under different Bounce-Bind parameters in the case of a 3-bit Ising problem. We obtain the probability distribution obtained by FPGA with Gibbs sampling at $\beta = 1$ and $10^5$ sweeps under no annealing mechanism. The blue bars stand for the probability distribution calculated by Non-Equilibrium Markov Chain algorithm mentioned in Supplementary Section S1 and orange bars stand for the data obtained from FPGA. The blue shaded bars in (i) are only for illustration due to the failure of the theory of random process, whose probability distribution is related to the initial state of the system.
• Figure 3: The basic spin architecture of Bounce-Bind Ising machine implemented by FPGA. We have respectively designed Ising machines based on the second-order Ising model (a) and third-order Ising model (b). $J$ and $h$ are only represented by 2 bits. In principle, they can also be represented by only one bit ($0$ for $-1$ and 1 for $+1$). When applied to the XORSAT problem, the bit width of $J$ and $h$ is 3 based on the Method \ref{['mtd1']}. For the third-order Ising machine, we mainly apply it to the 3R3X problem. In fact, it does not involve the second-order coupling term, so only the third-order coupling term is retained, and only two bits are used to represent $J$ and $h$. In (b), the brown multiplier represents the multiplication of two spins in the same clause. To reduce hardware cost, we used an XOR gate to implement this logic instead of using an $1\times1$ multiplier. Moreover, when $\mathcal{B}=0$, the Bounce-Bind spin unit reduces to the classical one.
• Figure 4: Success probability versus rounds taken for different Bounce-Bind parameters. We provide three examples to illustrate how the success probability varies with the number of samples under different Bounce-Bind parameters, where the black dashed line represents the maximum success probability at given sampling rounds. For optimal success probability, $\mathcal{B}$ needs to be carefully tuned for every size N and benchmark for the chosen sweeps. Data points are averaged over 100 random instances and 1 000 trials per instance, each trial taking the sweeps mentioned in the figure. Error bars on each of the graphs were generated by calculating the bootstrap resamples, with $95\%$ confidence interval for the parameter being estimated. In (b), we zoom in on the figure ranging from 0 to 5.
• Figure 5: Success probability versus Bounce-Bind parameter taken for different search rounds. For these three types of problems, we used one example each to illustrate how the success probability varies with the Bounce-Bind parameters across different sampling rounds. Data points are averaged over 100 random instances and 1 000 trials per instance, each trial taking the MCSs mentioned in the figure. Error bars on each of the graphs were generated by calculating the bootstrap resamples, with $95\%$ confidence interval for the parameter being estimated.