5ns per meter I think he said

There's been a lot of work by IBM and Intel on optics. Thunderbolt was supposed to be optical but the practicalities were that copper was easier to productionise at scale. There are network switches with photonics, search for Barefoot Tofino 2 co-packaged photonics

Nvidia made a point that Blackwell was able to keep using copper data lines. It sounded like it was difficult. The main problem with on package optical links, as I understand it, is that silicon is a bad light emitting material. This made it so it could never be a monolithic, single material design. Chiplets make it an option.

@@davidgunther8428 This is part of it. For context, I am a graduate student in optical sciences and work with integrated photonics regularly. Your comment is pretty accurate, but I feel like I should give a bit of extra context. Silicon is an excellent waveguide material, but it isn't an excellent material to make a light source (emitting material), as you said. The reason for this is that silicon is an indirect band gap semiconductor, so when you look at the conduction and valence band diagrams for it, the minimum of the conduction band is not in line with the maximum of the valence band. Essentially, in order for an electron to change from the conduction band state to the valence band state, it has to couple to a phonon in the silicon lattice (essentially a quantized lattice vibration). That means that the electron has to do two things, not just the one of emitting a photon. What this means practically is that silicon light sources are incredibly inefficient, which is exactly what you don't want with an integrated system. The electrons just don't want to make that transition quickly. Other semiconductors are direct band gap semiconductors, so an electron in an excited state can simply emit a photon and go on its merry way. Think gallium nitride and the other semiconductors used to make LEDs. What this means for an integrated system is that you'd have to make the emissive components on a chip of a different material and then directly couple that into silicon waveguides (again, silicon is an excellent material for this depending on wavelengths), which is just really expensive and complicated to do, so nobody does it. Making efficient waveguide lasers that can be used for integrated photonics is one of the biggest problems in the field right now, and there are a whole host of different possible solutions to the issue that are being tested. Silicon is an excellent guide material, particularly in the near infrared (where the telecom wavelengths are, like 1310nm and 1550nm). It has a very high refractive index (~3.5) and very low absorption at the same time. This is incredibly useful because higher refractive indices confine light tighter into waveguides, which lets you make tighter bends before you start losing light. Tighter bends means a smaller feature size and more density. Technically, germanium is a better material (it has an index of 4 and better transmission), but it is far more expensive, and we know how to work with silicon very well. So you might see companies putting special semiconductor laser diodes onto chips, or they might go for a different material. All of the photonics people I've talked to about this know that I/O will eventually be forced to go optical, but they're trying to stay electrical until they absolutely cannot anymore because optical is just really difficult and expensive at the moment. NVIDIA bragged about not needing to use optical transceivers in their new DGX racks, which are essentially devices that connect to the motherboard, not sending light directly to the chips. Needing all of those transceivers is quite the cost, so I don't blame them for sticking with copper. They are only sending signals a few meters anyway.

@@matthewmalaker477 thanks for an update on the state of the field. I've haven't been working in optoelectronics for a while.

@@matthewmalaker477Hi Matthew. When did Nvidia brag about not needing optical links in their new DGX racks? What is your source?

This has biggest benefits for supercomputers and such, so they will get it first. Only if the price comes down a lot will it be brought to consumers.

The benefit of optical is bigger on systems with larger physical scale. On smaller system the penalty from electric data transmission is much smaller due to shorter travel distance so I doubt it will come to consumer products anytime soon.

​@@hl321662I agree on that note, although if we take PCIe 5 for example, it starts to cost more because of signal integrity issues. If this continues to be a bigger problem with PCIe 6 then optical could be the more economical choice perhaps.