(Q|O)SFP are basically just raw high speed serial interfaces to whatever - you see this a lot in FPGAs, you can use the QSFP interfaces for anything high speed - PCIe, SATA, HDMI…
> Although we can already buy commercial transceiver solutions that allow us to use PCIe devices like GPUs outside of a PC, these use an encapsulating protocol like Thunderbolt rather than straight PCIe.
> [snip]
> As explained in the intro, this doesn’t come without a host of compatibility issues, least of all PCIe device detection, side-channel clocking and for PCIe Gen 3 its equalization training feature that falls flat if you try to send it over an SFP link.
So, uh… what’s the benefit? How much overhead does Thunderbolt really introduce, given it solves these other issues?
The benefits are twofold: physical colocation and bandwidth.
Thunderbolt 5 offers 80Gbps of bidirectional bandwidth. PCIe 5.0 16x offers 1024Gbps of bidirectional bandwidth. This matters.
TB5 cables can only get so long whereas fiber can go much farther more easily. This means that in a data center type environment, you could virtualize your GPUs and attach them as necessary, putting them in a separate bank (probably on the same rack).
The video is about a 2x1 link, which the author hopes to eventually scale up to 3x4 using 40 gig transceivers. I'd say thunderbolt is probably safe in the near future.
Bidirectional is a lot like biweekly. Biweekly depending on context means twice a week or once every two weeks and bidirectional can both mean per direction and total of both directions.
I'm only a single datapoint but I've never encountered that usage. My understanding of a bidirectional link is that it meets the same spec in both directions simultaneously. It's important precisely because many links aren't bidirectional, sharing a single physical link between two logical links.
This was a super interesting video to watch. I honestly thought SFP required more setup, but this explains why AliExpress is so ripe with USB3 and HDMI over SFP converters that are dirt cheap.
So you're saying I can put a handful of 4090's out in the middle of snowy Michigan with a handful of OM4 cables snaking into my basement to run legit arctic cooling with no noise?
While at a higher level, thunderbolt and https://en.wikipedia.org/wiki/ExpEther can both of course work over fiber too!
(Q|O)SFP are basically just raw high speed serial interfaces to whatever - you see this a lot in FPGAs, you can use the QSFP interfaces for anything high speed - PCIe, SATA, HDMI…
> [snip]
> As explained in the intro, this doesn’t come without a host of compatibility issues, least of all PCIe device detection, side-channel clocking and for PCIe Gen 3 its equalization training feature that falls flat if you try to send it over an SFP link.
So, uh… what’s the benefit? How much overhead does Thunderbolt really introduce, given it solves these other issues?
Thunderbolt 5 offers 80Gbps of bidirectional bandwidth. PCIe 5.0 16x offers 1024Gbps of bidirectional bandwidth. This matters.
TB5 cables can only get so long whereas fiber can go much farther more easily. This means that in a data center type environment, you could virtualize your GPUs and attach them as necessary, putting them in a separate bank (probably on the same rack).
Good luck getting a 1Tbit tranceiver. Anydirectional. Also it's 512Gbitish per direction.
But yes I meant 512Gbps each way, to be clear.
There is an interesting NSDI talk on the paper too - https://www.youtube.com/watch?v=kDJHA7TNtDk (2023)