The proposed plan by the corporate is to make use of an optical NVLink equal. While the present NVLink 2.Zero chip makes use of eight pico Joules per bit (Eight pJ/b) and may ship alerts solely to 0.three meters with none repeaters, the optical substitute is able to sending information wherever from 20 to 100 meters whereas consuming half the ability (four pJ/b). NVIDIA has conceptualized a system with 4 GPUs in a tray, all of that are related by gentle. To energy such a setup, there are lasers that produce 8-10 wavelengths. These wavelengths are modulated onto this at a pace of 25 Gbit/s per wavelength, utilizing ring resonators. On the receiving aspect, ring photodetectors are used to choose up the wavelength and ship it to the photodetector. This approach ensures quick information switch able to lengthy distances.
And that isn’t even all of it. To combine every thing, NVIDIA wants to mix it with chip packaging expertise. Pictured beneath is a setup the place the GPU and reminiscence mixture sits on a silicon interposer, which itself is positioned on prime of an natural bundle. This natural bundle is related to the photonic built-in circuit (PIC). Further, there may be {an electrical} interface chip (EIC) that converts {the electrical} alerts coming from the GPU, utilizing resonators, to optical alerts. Optical interconnect wants a equally constructed change on its aspect.
Using these strategies, NVIDIA says that sooner or later there might be techniques with 1000s of interconnected GPUs, every of which talks to a different utilizing a pace of sunshine. It is just a matter of time earlier than the expertise reaches the market, nonetheless, there are not any ensures as that is solely a analysis mission for now.
