Google Apollo: The >$3B Game-Changer in Datacenter Networking

semianalysis.com

45 pointsalphabetting5 comments

As you can get higher radix fix becomes more interesting IMO. Less about switching who you talk to (although you do some of that) and more about how many channels you allocate.

If future optical switches (or this one?) are multi-wavelength capable, the radix goes up! Even just separating a high & low band on each transceiver, making multiple links out of it (would require more complex transceivers) would be super useful/cool.

I wonder how much Segment-Routing (SR) this all uses, an IP feature where packets pre-enscribe their route- or whether it's all custom stuff. Seems like a useful place for SR.

They use some of this too btw. Also wavelength level routing happens with breakout cables from ToR to Compute.

Isn't there a certain sadness that all this clever technology is in the service of collecting info for ads?

I might try to phrase it another way: isn't it sad that so few other enterprises have the resources & scale to try for similar? Why are there so few others in the world who might try these things?

Ads doesn't need this. Ads pay for this.

Reply on news.ycombinator.com