Ultra-scalable high-performance computing (HPC) network using dense wavelength-division multiplexing (DWDM)

A system, comprising:

• a first computing system, comprising:
• a plurality of graphics processing unit (“GPU”) devices, each GPU device comprising one or more GPUs and an optoelectronic device;
• a plurality of GPU multiplexer/demultiplexer (“mux/demux”) devices, each GPU mux/demux device being communicatively coupled with a first number of optoelectronic devices of a corresponding first number of GPU devices;
• a plurality of optical circuit switches (“OCSs”), each OCS comprising a second number of input/output (“I/O”) ports and a third number of micro-electromechanical systems (“MEMS”) mirrors, each MEMS mirror being configured to selectively route optical signals between one I/O port and another I/O port among the second number of I/O ports;
• a plurality of wavelength selective switches (“WSSs”), each WSS comprising a plurality of WSS mux/demux devices, a fourth number of WSS mux/demux devices each being communicatively coupled with a GPU mux/demux device among the plurality of GPU mux/demux devices and a fifth number of WSS mux/demux devices each being communicatively coupled with an I/O port among the second number of I/O ports of an OCS among the plurality of OCSs, each WSS mux/demux device being configured to communicatively couple with each of two or more other WSS mux/demux devices among the plurality of mux/demux device, wherein each WSS is configured to selectively route optical signals between two of the WSS mux/demux devices;
• a plurality of amplifiers, each amplifier being configured to amplify an optical signal between a GPU mux/demux device among the plurality of GPU mux/demux devices and a WSS among the plurality of WSSs;
• and
• a second computing system configured to control the plurality of OCSs and the plurality of WSSs to selectively route optical signals between two of the WSS mux/demux devices of one or more WSSs and to selectively route optical signals between I/O ports of one or more OCSs to connect a sixth number of GPU devices in a first network topology among a plurality of network topologies for computing a series of computations.

A computer-implemented method, comprising:

• receiving first input indicating a first network topology among a plurality of network topologies to connect a plurality of graphics processing unit (“GPU”) devices of a system, each GPU device comprising one or more GPUs and an optoelectronic device, at least one GPU device among the plurality of GPU devices further comprising one or more GPU I/O ports, the system further comprising:
• a plurality of GPU multiplexer/demultiplexer (“mux/demux”) devices, each GPU mux/demux device being communicatively coupled with a first number of optoelectronic devices of a corresponding first number of GPU devices;
• a plurality of optical circuit switches (“OCSs”), each OCS comprising a second number of input/output (“I/O”) ports and a third number of micro-electromechanical systems (“MEMS”) mirrors, each MEMS mirror being configured to selectively route optical signals between one I/O port and another I/O port among the second number of I/O ports;
• and
• a plurality of wavelength selective switches (“WSSs”), each WSS comprising a plurality of WSS mux/demux devices, a fourth number of WSS mux/demux devices each being communicatively coupled with a GPU mux/demux device among the plurality of GPU mux/demux devices and a fifth number of WSS mux/demux devices each being communicatively coupled with an I/O port among the second number of I/O ports of an OCS among the plurality of OCSs, each WSS mux/demux device being configured to communicatively couple with each of two or more other WSS mux/demux devices among the plurality of mux/demux device, wherein each WSS is configured to selectively route optical signals between two of the WSS mux/demux devices;
• controlling the plurality of WSSs and the plurality of OCSs to selectively route optical signals between the two of the WSS mux/demux devices of one or more WSSs and to selectively route optical signals between I/O ports of one or more OCSs to connect a sixth number of GPU devices in the first network topology for computing a series of computations;
• in response to receiving, via at least one GPU I/O port among the one or more GPU I/O ports, second input comprising at least one of input data or input prompts, performing the series of computations over the first network topology using the at least one of input data or input prompts;
• and outputting, via the at least one GPU I/O port, results of the series of computations.