Comments on: One Laser To Pump Up AI Interconnect Bandwidth By 10X

By: Calamity Jim

Calamity Jim — Fri, 18 Oct 2024 11:39:49 +0000

In reply to Slim Albert. Right on! And don't forget those wavelengths come out of just one laser, for some quite cost- and space-effective secret sauce, of interconnect gastronomy!

By: Timothy Prickett Morgan

Timothy Prickett Morgan — Thu, 17 Oct 2024 13:15:14 +0000

In reply to Eric Olson. Could not agree more, and there are others who are going to push this interconnect hard. I am taking a hard look at Celestial AI now as well, and a few others that are on the periphery. This is only just starting to get interesting.

By: Slim Albert

Slim Albert — Thu, 17 Oct 2024 10:27:42 +0000

Impressive — Xscape photonics’ ChromX laser, with its 128+ independently modulated wavelengths is just awesome! It’s great that spacing (between wavelengths) can be configured, wider to account for phase shifts and dispersion in longer runs of fiber, and narrower (with more multiplexed wavelengths) for incresed bandwidth in shorter ones. I imagine that some of the signalling/wavelengths would be used for ECC and clocking (sync) so an 80-wavelength configuration might transmit a 64-bit data payload per clock, through a single fiber, which is quite impressive. This puts the “parallel” back in SerDes (in a big way)!

By: Eric Olson

Eric Olson — Thu, 17 Oct 2024 05:00:42 +0000

Could network-bandwidth bottlenecks imply the low-powered H20 GPUs designed for the Chinese market lead to more efficient AI training clusters?

On the other hand, maybe the topology of the neural network could be adapted to the topology of network hardware so nothing is bottlenecked. I suspect careful co-design of the neural network and the network hardware over the next five years will yield even greater performance increases than using a nuclear reactor to power the training cluster.