Periodic Reporting for period 1 - SWIM (A SpinWave Ising Machine)
Période du rapport: 2023-07-01 au 2025-06-30
Rapid and energy-efficient solvers of combinatorial optimization problems are necessary in numerous applications molecular assembly and protein folding in medicine, high frequency trading in finance and circuit design optimization in electronics. Ising machines (IMs) are physics-inspired combinatorial optimization solvers which can accelerate solving a wide class of problems - nondeterministic polynomial time (NP)-hard and NP-complete problems with a time-to-solution parameter. Ising machines improve time-to-solution scaling for combinatorial problems from exponential to sub-exponential or even polynomial scaling for certain subclass of problems. The SWIM project tackles a critical challenge for the emerging field of Ising Machines: the development of miniaturized, low-power, and low-cost IMs with high computational power, commercial feasibility, and speed beyond classical Von-Neuman and CMOS-based solutions.
The SWIM project is built upon the achievements of Coherent Ising machines which use optical pulses as equivalent Ising spins. CIMs offer thousands of computational spins for large combinatorial problems but are bulky, highly consuming and temperature unstable. The SWIM project solves these disadvantages through the use of exceptionally slow and low-power propagating spinwaves in an novel architecture of time-multiplexed spinwave Ising Machines. Because spin-waves propagate more than five orders of magnitude slower than light, an entire time-multiplexed Ising network fits on a compact mm-size yttrium-iron-garnet (YIG) chip and is driven with standard ready-from-the-shelf microwave electronics. This solutions reduces hardware volume, slashes power consumption from kilowatts to milliwatts, and paves the way for chip integration.
The SWIM project is built upon the achievements of Coherent Ising machines which use optical pulses as equivalent Ising spins. CIMs offer thousands of computational spins for large combinatorial problems but are bulky, highly consuming and temperature unstable. The SWIM project solves these disadvantages through the use of exceptionally slow and low-power propagating spinwaves in an novel architecture of time-multiplexed spinwave Ising Machines. Because spin-waves propagate more than five orders of magnitude slower than light, an entire time-multiplexed Ising network fits on a compact mm-size yttrium-iron-garnet (YIG) chip and is driven with standard ready-from-the-shelf microwave electronics. This solutions reduces hardware volume, slashes power consumption from kilowatts to milliwatts, and paves the way for chip integration.
In the course of the project two time-multiplexed Ising-machine (IM) concepts were developed, preserving the strengths of Coherent Ising Machines (dense connectivity, fast time-to-solution parameter) while key limitations to miniaturisation and thermal stability were removed. The first IM was designed with propagating spinwaves. For this, a ring oscillator was built, in which artificial spins are represented with spin-wave RF pulses propagating in a 10-mm-long 5-μm-thick Yttrium Iron Garnet (YIG) film and having the phases 0 and 180 degrees relative to the reference signal. A ring oscillator comprises a phase-sensitive amplifier which binarises the phases. The couplings are implemented by delayed pulse injection, following the CIM time-multiplexing principle. The prototype supports an 8-spin MAX-CUT instance and finds solutions in less than 4 μs with approximately 7 μJ energy per run. This proof-of-concept established the basic operation, quantified the dispersion-limited spin capacity of a single track, and outlined a CMOS-compatible scaling path (hundreds to ~10^5 spins via multi-track waveguides and integrated amplification).
To overcome the spin-wave dispersion and limits of YIG, surface acoustic wave (SAW)-based ring oscillator with intrinsically linear dispersion and high thermal stability was developed. The system implements a fully programmable, all-to-all 50-spin Ising machine using a commercially available SAW delay line, microwave phase-sensitive amplification, and an FPGA measurement-and-feedback block analogous to that used in state-of-the-art CIMs. Single-run compute time is 10 ms; total power is 1.82 W; energy per solution is 18.2 mJ; and the figure of merit reaches 55 solutions/s/W. Thanks to operation at 320 MHz and short acoustic delay line the thermal stability is improved by 4–5 orders of magnitude relative to optical CIMs and by 1–2 orders relative to SWIM, enabling stable room-temperature operation without precision thermostats or phase locked loop (PLL) systems.
For system characterization, a comprehensive benchmarking was performed with 170 random 50-spin MAX-CUT instances (BiqMac) across densities 0.1–0.9 with 500 runs per instance. For 99%-accurate solutions, success probabilities are high, with an optimum-coupling operating point yielding approximately 84% average success at density 0.5 and reducing the 99% time-to-solution from 597 ms to 25 ms. Exact-solution rates are comparable to reported 100-spin CIM behaviour, with notably stronger performance at high graph density.
To overcome the spin-wave dispersion and limits of YIG, surface acoustic wave (SAW)-based ring oscillator with intrinsically linear dispersion and high thermal stability was developed. The system implements a fully programmable, all-to-all 50-spin Ising machine using a commercially available SAW delay line, microwave phase-sensitive amplification, and an FPGA measurement-and-feedback block analogous to that used in state-of-the-art CIMs. Single-run compute time is 10 ms; total power is 1.82 W; energy per solution is 18.2 mJ; and the figure of merit reaches 55 solutions/s/W. Thanks to operation at 320 MHz and short acoustic delay line the thermal stability is improved by 4–5 orders of magnitude relative to optical CIMs and by 1–2 orders relative to SWIM, enabling stable room-temperature operation without precision thermostats or phase locked loop (PLL) systems.
For system characterization, a comprehensive benchmarking was performed with 170 random 50-spin MAX-CUT instances (BiqMac) across densities 0.1–0.9 with 500 runs per instance. For 99%-accurate solutions, success probabilities are high, with an optimum-coupling operating point yielding approximately 84% average success at density 0.5 and reducing the 99% time-to-solution from 597 ms to 25 ms. Exact-solution rates are comparable to reported 100-spin CIM behaviour, with notably stronger performance at high graph density.
For the first time a spinwave based time-multiplexed Ising Machine (SWIM) has been demonstrated. Spinwave dispersion was identified as a main limitation for increasing the number of spins. To overcome this limitation, a surface-acoustic-wave (SAW)-based Ising machine was developed supporting 50 spins and demonstrating much simpler scalability pathway. The results and the projections towards future SWIM and SAWIM have been published in the journal Communications Physics.
The project has lead to an entirely new direction within spintronics, magnonics, acoustics, and time-multiplexed Ising Machines and the results have been presented at several international conference with great interest from the research field. Hence, the impact from the SWIM project is expected to last for a long time, hopefully also leading to a new commercially viable Ising Machine technology platform.
The project has lead to an entirely new direction within spintronics, magnonics, acoustics, and time-multiplexed Ising Machines and the results have been presented at several international conference with great interest from the research field. Hence, the impact from the SWIM project is expected to last for a long time, hopefully also leading to a new commercially viable Ising Machine technology platform.