Periodic Reporting for period 1 - SPINTOP (SPINTOP)
Période du rapport: 2022-04-01 au 2023-09-30
There exists a class of important computational problems that conventional computers are unable to address with reasonable efficiency. Such Combinatorial Optimization (CO) problems are pervasive in a wide range of critically important sectors of society, e.g. in business operations, manufacturing, and research, including man-power scheduling, vehicle routing, IC circuit layout, protein folding and DNA sequencing, efficient big-data clustering, election modelling, network diagnosis, modelling molecular dynamics, discovery of new medicines/chemicals/materials, and so forth. At present, the CO market size is of the order of €1B and is expected to exhibit a 56% market growth rate with a 2030 market size forecast at €50B.
Since conventional computers are ineffective in handling large CO problems, dedicated hardware – both quantum and quantum-inspired – are intensely researched and developed world-wide. While quantum computers currently receive most of the attention, they all face essentially insurmountable challenges in the near-term perspective. Quantum-inspired alternative – so-called Ising Machines – have been developed over the last 20 years by D-wave in superconducting technology and is commercially available. Various other types of Ising machines have been proposed, exploiting novel physical building blocks such as spintronic devices, memristor crossbars, metal-insulator relaxation oscillators, and degenerate optical parametric oscillators, as well as conventional CMOS technology with analog electric oscillators, and field programmable gate arrays (FPGAs). In particular, optical Coherent Ising Machines (CIM) have attracted great attention due to their high computational speed, a time-multiplexing method that provides all-to-all Ising spin connections, and the largest amount of supported Ising spins amongst all implementations. Nevertheless, the commercial feasibility of optical CIMs remains elusive as the technology requires optical tables, kilowatts of power, and kilometers of optical fibers, which altogether blocks its further development from a proof-of-principle demonstration to a miniaturized commercially viable device.
In SPINTOP, we develop a novel time-multiplexed spinwave Ising machine (SWIM) with artificial spin states implemented via the phase of spinwave radio-frequency (RF) pulses propagating in an Yttrium Iron Garnet (YIG) thin film. It is in principle similar to the optical CIM but operates at orders of magnitude lower frequencies in the microwave domain, which greatly simplifies the peripheral eletronics, and instead of optical pulses uses spin wave pulses, which makes it possible to miniaturize the Ising Machine by orders of magnitude.
Since conventional computers are ineffective in handling large CO problems, dedicated hardware – both quantum and quantum-inspired – are intensely researched and developed world-wide. While quantum computers currently receive most of the attention, they all face essentially insurmountable challenges in the near-term perspective. Quantum-inspired alternative – so-called Ising Machines – have been developed over the last 20 years by D-wave in superconducting technology and is commercially available. Various other types of Ising machines have been proposed, exploiting novel physical building blocks such as spintronic devices, memristor crossbars, metal-insulator relaxation oscillators, and degenerate optical parametric oscillators, as well as conventional CMOS technology with analog electric oscillators, and field programmable gate arrays (FPGAs). In particular, optical Coherent Ising Machines (CIM) have attracted great attention due to their high computational speed, a time-multiplexing method that provides all-to-all Ising spin connections, and the largest amount of supported Ising spins amongst all implementations. Nevertheless, the commercial feasibility of optical CIMs remains elusive as the technology requires optical tables, kilowatts of power, and kilometers of optical fibers, which altogether blocks its further development from a proof-of-principle demonstration to a miniaturized commercially viable device.
In SPINTOP, we develop a novel time-multiplexed spinwave Ising machine (SWIM) with artificial spin states implemented via the phase of spinwave radio-frequency (RF) pulses propagating in an Yttrium Iron Garnet (YIG) thin film. It is in principle similar to the optical CIM but operates at orders of magnitude lower frequencies in the microwave domain, which greatly simplifies the peripheral eletronics, and instead of optical pulses uses spin wave pulses, which makes it possible to miniaturize the Ising Machine by orders of magnitude.
We have succeeded in developing a spinwave based time-multiplexed Ising Machine (SWIM) with up to 10 spins, and used it to solve MAX-CUT problems with up to 8 spins. The SWIM is implemented using a 10-mm-long 5-μm-thick Yttrium Iron Garnet film with off-the-shelf microwave components and solves an 8-spin MAX-CUT problem in less than 4 μs consuming only 7 μJ.
Although the number of spins is modest, the demonstration is the first of its kind. The presented SWIM was implemented with μm-thick YIG waveguides having mm-scale lateral dimension. However, in the recent decade, magnonics has made significant progress in scalability demonstrating nanometer-thick YIG-based waveguides and signal processing devices with sub-micrometer channel width. At the nanoscale, the spinwave bandwidth exceeds hundreds of MHz and the delay deviation can be improved with the same methods applicable to mm-scale YIG films. The group velocity of spinwaves in the magnetostatic range is inversely proportional to the thickness of the waveguides and can reach tens of meters per second. This makes it beneficial to exploit nm-thick YIG films with reduced SW velocity to downscale the YIG delay line in terms of physical size. We have estimated the maximum number of spins to just below 800 that can realistically fit and propagate without too much loss or dispersion in a YIG based delay line. In order to further improve the spin capacity of the proposed SWIM, an array of microscale YIG waveguides should be used to form multiple tracks, each with its own linear and phase-sensitive amplifiers. In a multi-track system, the interconnection between propagating spins can be implemented similarly to the optical CIM with digital FPGA or application-specific integrated circuit (ASIC) blocks. In the case of using N parallel tracks instead of connecting all the delay lines and amplifiers in series in the time-multiplexed architecture of SWIM, the time-to-solution parameter will be improved and reduced by N since the circulation time will be N times shorter and the propagating RF pulses will interact with each other more frequently.
We hence suggest an array of 150 parallel waveguides with a lateral size of 2660 × 1.0 μm each supporting about 800 spins. A separation of 10 μm between waveguide centers would make cross-talk negligible. Then, with the total lateral size of 2.66 × 1.5 mm2, a multi-track SWIM would have 120,000 spins, i.e. about the same capacity of the most recent optical CIM. However, whereas the optical CIM requires a 5-km thermally-stabilized optical fiber delay system, our future SWIM would be realized in a singla film with the dimensions 2.66 × 1.5 mm2.
Although the number of spins is modest, the demonstration is the first of its kind. The presented SWIM was implemented with μm-thick YIG waveguides having mm-scale lateral dimension. However, in the recent decade, magnonics has made significant progress in scalability demonstrating nanometer-thick YIG-based waveguides and signal processing devices with sub-micrometer channel width. At the nanoscale, the spinwave bandwidth exceeds hundreds of MHz and the delay deviation can be improved with the same methods applicable to mm-scale YIG films. The group velocity of spinwaves in the magnetostatic range is inversely proportional to the thickness of the waveguides and can reach tens of meters per second. This makes it beneficial to exploit nm-thick YIG films with reduced SW velocity to downscale the YIG delay line in terms of physical size. We have estimated the maximum number of spins to just below 800 that can realistically fit and propagate without too much loss or dispersion in a YIG based delay line. In order to further improve the spin capacity of the proposed SWIM, an array of microscale YIG waveguides should be used to form multiple tracks, each with its own linear and phase-sensitive amplifiers. In a multi-track system, the interconnection between propagating spins can be implemented similarly to the optical CIM with digital FPGA or application-specific integrated circuit (ASIC) blocks. In the case of using N parallel tracks instead of connecting all the delay lines and amplifiers in series in the time-multiplexed architecture of SWIM, the time-to-solution parameter will be improved and reduced by N since the circulation time will be N times shorter and the propagating RF pulses will interact with each other more frequently.
We hence suggest an array of 150 parallel waveguides with a lateral size of 2660 × 1.0 μm each supporting about 800 spins. A separation of 10 μm between waveguide centers would make cross-talk negligible. Then, with the total lateral size of 2.66 × 1.5 mm2, a multi-track SWIM would have 120,000 spins, i.e. about the same capacity of the most recent optical CIM. However, whereas the optical CIM requires a 5-km thermally-stabilized optical fiber delay system, our future SWIM would be realized in a singla film with the dimensions 2.66 × 1.5 mm2.
For the first time a spinwave based time-multiplexed Ising Machine (SWIM) has been demonstrated. The results and the projections towards future SWIM have been published in the journal Communications Physics. Four patents have also been filed, a start-up company SpinWave Computing has been launched and discussions with several Swedish Venture Capital firms have been initiated.
To scale the SWIM to much larger number of spins, more research and development will be needed. Development of FPGA based external control of SWIMs has begun and will need to continue in order to demonstrate SWIMs with much higher number of spins.
The project has lead to an entirely new direction within spintronics, magnonics, and time-multiplexed Ising Machines and the results have been presented at several international conference with great interest from the research field. We hence believe the impact from the SPINTOP project will be significant and last for a long time, hopefully also leading to a new commercially viable Ising Machine technology platform.
To scale the SWIM to much larger number of spins, more research and development will be needed. Development of FPGA based external control of SWIMs has begun and will need to continue in order to demonstrate SWIMs with much higher number of spins.
The project has lead to an entirely new direction within spintronics, magnonics, and time-multiplexed Ising Machines and the results have been presented at several international conference with great interest from the research field. We hence believe the impact from the SPINTOP project will be significant and last for a long time, hopefully also leading to a new commercially viable Ising Machine technology platform.