(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?
I go over it in the video but yes, active thunderbolt is probably a very good choice for a lot of people. I went into another direction for some reasons that are not applicable to everyone:
- Learning : I want to learn about the lower level of PCIe and it's a good project.
- Re-use of cabling : I have a bunch of single mode fiber bundle going around already. You can't find thunderbolt that just have a LC connector ...
- Isolation : Active thunderbolt cable still often have copper for some low speed signals, they don't offer true galvanic isolation
- Avoid dealing with thunderbolt. I want a custom chassis/pcb at one end and chips to convert from TB back to PCIe are not readily available to make custom stuff with ... (not as an individual anyway).
So yeah, if you want a ready to use solution, TB cable is absolutely a good choice, here I'm having some fun, learning in the process and hopefully sharing some of the knowledge.
Hey, I love a great self-educational deep dive. Don’t have time to watch the video until after the workday, but it sounds enlightening! (I swear that was not intentional.)
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).
Active optical (yes!) Thunderbolt cables can be much longer. After all, optical fiber was the original medium for Thunderbolt, back when it was still called Light Peak.
As for bandwidth, the medium transition seems to actually limit the author’s capabilities by losing some of the more advanced link-training features that are necessary for the highest-bandwidth PCIe 3 connections, never mind PCIe 5.
Hundreds of meters is considered short range in the world of *SFP. If you just plan on putting the GPUs in the same rack then I'm not sure it really matters, but you can really put anything anywhere in your DC and have things zoned with *SFP.
I don't think there is any reason TB couldn't do the same, beyond it would be even more niche to want non-modular/patchable cables+transceivers at those lengths (especially since fiber is often bundled dozens/hundreds of strands over a single trunk cable between racks).
Thunderbolt is kind of cursed. To insure maximum compatibility it mandates a legacy usb2 connection via separate connections. TB3/USB4/TB4 are packetized, but afaik there's no defined way to packetized usb2, it's expected there be a whole separate set of wires for it.
And because of timings, my admittedly so-aonunderstanding that you can only get about 7m before you absolutely have to have a hub/repeater (unless you can speed up the speed of light considerably). This limit to how long a single length can be can't really be cheated without violating usb specs.
It's awesome if folks have packetized USB2. A pity it's not in the flipping spec though!!
That Corning made it 50m is wild. You need a virtual hub at the start that can pretend to be hubs 1-5 (so it's close enough to time well). Then a hub on the other side of the cable at (skinny) tree depth 6. Allowing for 4 devices under it (the number of ports on a usb2 hub in the spec. But you could work around by faking being not a skinny tree but a fat tree, maybe?).
IIRC, USB-PD also requires USB 2.0 signaling. The idea of dedicated lower-bandwidth signaling wires isn’t uncommon in my very limited EE experience—level 3 charging reuses J-1772 signaling to control the charge available of the DC pins.
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.
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.
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.
Worth noting too that well respected vendors have been selling optical thunderbolt cables for a while now. I wonder if they are length limited for latency reasons (& hello hollow core fiber)? I wonder if they are usb3/multiprotocol, or if they are usb4 only. I also wonder how they handle the incredibly jank usb4 requirement to also have a separate legacy usb2. As a usb-c enjoyer, I can still admit: sure seems like USB is a lot of work to support! I can't help but wonder how blissfully simple a future CXL over cable stack might look by compare.
https://www.owc.com/solutions/usb4-cables
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?
39 comments
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…
> 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?
- Learning : I want to learn about the lower level of PCIe and it's a good project. - Re-use of cabling : I have a bunch of single mode fiber bundle going around already. You can't find thunderbolt that just have a LC connector ... - Isolation : Active thunderbolt cable still often have copper for some low speed signals, they don't offer true galvanic isolation - Avoid dealing with thunderbolt. I want a custom chassis/pcb at one end and chips to convert from TB back to PCIe are not readily available to make custom stuff with ... (not as an individual anyway).
So yeah, if you want a ready to use solution, TB cable is absolutely a good choice, here I'm having some fun, learning in the process and hopefully sharing some of the knowledge.
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).
I couldn’t find any optical TB5 cables, but here’s a 4.5m TB4 one: https://www.owc.com/blog/the-new-superlong-40gb-s-owc-active...
And if TB3 is enough, Corning makes them in lengths up to 50m: https://www.corning.com/microsites/coc/oem/documents/ocbc/OE...
As for bandwidth, the medium transition seems to actually limit the author’s capabilities by losing some of the more advanced link-training features that are necessary for the highest-bandwidth PCIe 3 connections, never mind PCIe 5.
I don't think there is any reason TB couldn't do the same, beyond it would be even more niche to want non-modular/patchable cables+transceivers at those lengths (especially since fiber is often bundled dozens/hundreds of strands over a single trunk cable between racks).
And because of timings, my admittedly so-aonunderstanding that you can only get about 7m before you absolutely have to have a hub/repeater (unless you can speed up the speed of light considerably). This limit to how long a single length can be can't really be cheated without violating usb specs.
It's awesome if folks have packetized USB2. A pity it's not in the flipping spec though!!
That Corning made it 50m is wild. You need a virtual hub at the start that can pretend to be hubs 1-5 (so it's close enough to time well). Then a hub on the other side of the cable at (skinny) tree depth 6. Allowing for 4 devices under it (the number of ports on a usb2 hub in the spec. But you could work around by faking being not a skinny tree but a fat tree, maybe?).
> 1024Gbps
Good luck getting a 1Tbit tranceiver. Anydirectional. Also it's 512Gbitish per direction.
There's already 800Gb transceivers readily available, 1.6 is probably getting preview deploys to some hyperscalers & other early adopters as we speak.
But yes I meant 512Gbps each way, to be clear.
Worth noting too that well respected vendors have been selling optical thunderbolt cables for a while now. I wonder if they are length limited for latency reasons (& hello hollow core fiber)? I wonder if they are usb3/multiprotocol, or if they are usb4 only. I also wonder how they handle the incredibly jank usb4 requirement to also have a separate legacy usb2. As a usb-c enjoyer, I can still admit: sure seems like USB is a lot of work to support! I can't help but wonder how blissfully simple a future CXL over cable stack might look by compare. https://www.owc.com/solutions/usb4-cables