UbiSwitch + Jetson Orin AGX Industrial: Connecting XFI to SFI/10GBASE-R/USXGMII

I’m currently redesigning the electronics of our subsea robots (ROV). We already use an UbiSwitch with baseboard on our surface unit, and I would like to integrate the Ubswitch module directly onto my PCB inside the ROV (in the electronics pod, with is in air at 0.8 bar).

I’m following the design guide (https://botblox.io/content/BB-AN-0004 - Designing a custom board for UbiSwitch.pdf) to design my PCB.
The design guide suggests using directly a PHY (instead of an SFP+ module) to get 10Gbps copper link, which I would like to achieve in order to communicate with our Jetson Orin AGX Industrial. The design guide asks to contact you for how to do.

Could you please advise me how best to get a 10Gbps link between the port 0 (or 9 or 10 in it’s easier) of the UbiSwitch and the Jetson Orin AGX Industrial?

Some information about the Jetson Orin AGX Industrial :

  • it has only a MAC, so the PHYs can be chose on both sides
  • the Jetson supports XFI (and the hardware supports SFI, but Nvdia don’t provides any support for SFI, so excepted if you have already done it and can share how to do it, we will tick to XFI)
  • the Jetson will be mounted on the same PCB as the Ubiswitch
  • distance between the Jetson and the ubiswitch can be around 5-8cm if needed, supporting longuer distances is a nice to have (our PCB will be about 40cm long, and it will simplify routing if I can place the ubiswitch farther away)

And 2 other smaller questions :

  1. for how long is the UbiSwitch Module garanteed to be available? (there is a mention of 5 years garantee from first release, but I couln’t find the release date)? How long will be the delay between anouncement of end of life and last time to buy?
  2. Is there any requirement on the inductance of the magnetics for the 1Gbps PHYs? can I use any magnetics rated for 1Gbps? (I’m planning to do PoE++ injection, so of course, I will take some rated for PoE++)

It would be simpler to connect the Jetson and UbiSwitch directly over XFI. Placing PHYs on either side just to go 5-8cm doesn’t make much sense. It adds design complexity and increases power draw. Are you sue the Jetson’s software doesn’t support XFI, do you have any documentation on that you can share with me?

For your other questions.

1) for how long is the UbiSwitch Module garanteed to be available? (there is a mention of 5 years garantee from first release, but I couln’t find the release date)? How long will be the delay between anouncement of end of life and last time to buy?

Go to the “Revision Tracker” tab on the product page of UbiSwitch, it shows the dates of release of each revision. In this case 5 Feb 2023. The 5 year lifetime is either renewed or not at the start of each year. It was renewed at the start of 2025 so the lifetime is to 2030. We plan to continue renewing this every year, as this module is used by many customers. The delay between announcement and last time buy will be 5 years.

**
2) Is there any requirement on the inductance of the magnetics for the 1Gbps PHYs? can I use any magnetics rated for 1Gbps? (I’m planning to do PoE++ injection, so of course, I will take some rated for PoE++)**

350uH is the required inductance for 1GBASE-T.

Thanks for your answers!

Is the UbiSwitch capable of XFI? (if so it’s great news, but I haven’t seen it listed in the datasheet). Or is it a typo and you mean SFI? If direct connection without a PHY is possible, it would of course be ideal!

I join the design guide for the Jetson Orin AGX familly (also available from Nvidia’s website, but requires signing up) and the datasheet.
XFI is the recommend way to do 10Gbps ethernet on the Jetson Orin AGX.
For SFI, there are various hints that it’s not supported correctly :

The same for USXGMII (that is only mentionnent in the datasheet, not even in the design guide) :

For the 2 other questions :
great if there is a 5 years delay between end of life anouncement and end of production.
I guessed it would be the classial 350µH, thanks for confirming

UbiSwitch can support the following modes on the MAC ports. This is done via the Command Line Interface on the UART port. More details on using that port is here.

So UbiSwitch supports USXGMII. It does not support XFI, that was a typo. For SFI, then you would select 10gbaser, however that’s designed for connecting to 10G SFPs, I actually don’t know if it would work with the Jetson. Unfortunately UbiSwitch does not support XFI, below is what each port does support.

Looking through Nvidia’s answers, you’re right it seems they only support XFI in software. That is quite restrictive because then, how would you even put a PHY on the 10G port of the Jetson? If it only supports XFI then you could only support an XFP transceiver, unless I misunderstand?

Thanks for confirming.

For the Jetson, it seems there are a few 10G base T PHYs that support XFI (they use a Marvell one on the devkit, and I found 2 from Broadcom), but so far, I found none that is available from reliable distributors and provides the dataset without having to sign a NDA first. So getting a 10Gbps PHY on the Jetson is possible, ut not as easy as I thought.

I also have a spare PCIe bus on the Jetson (PCIe x8, Gen 4). Do you have any experience on integrating the botblox at 10Gbps through PCIe? (converting from PCIe to SFI or USXGMII ?)

I see. Yes it is generally the way with the 10G PHYs that they’re difficult to get access to and support is patchy.

The MAC ports on UbiSwitch support 1000BASE-X, SGMII, 2500BASE-X, 5GBASE-R, 10GBASE-R (SFI), USXGMII. These can be configured over the Command Line Interface. Do you know if the Jetson can support any of these modes?

If not then I would guess the only way is to convert both sides to 10GBASE-T. We are considering developing a stackable USXGMII to 10GBASE-T that can be used with UbiSwitch. This would contain all the firmware and electronics necessary to achieve the conversion, on a stackable board. Would this be of interest?

Regarding PCIe to SFI for UbiSwitch, we haven’t seen this done before.

In hardware, the Jetson supports XFI, SFI and USXGMII. However, Nvidia refuses to support SFI and USXGMII in software, so I haven’t heard of anyone successfully using those. So I fear I’m stuck with XFI (or PCIe).

NB : there is also a 1Gbps RGMII port on the Jetson that I plan to use as a backup solution, but getting 10Gbps would be really nice.

For the stakable USXGMII to 10GBaseT, it really depends how it is done : I still need to get 2 SFP+ cages for redundant optical fiber uplink, and to get the 8 PHYs onto my PCB (for routing to some devices, and doing PoE+/++ injection on some others). So it will only do the job if it converts a single 10Gbps MAC into 10GBase-T, and pass through all other signals (which is probably too much a niche application for you to do a stackable board).

So I fear I will have to go the “heavy” way, and convert to 10GBASE-T on both sides.

Are there any 10GBASE-T PHYs that you have validated to work well with the Ubiswitch (without needing an SFP+ cage+module)?

Honestly, I wouldn’t recommend doing a 10GBASE-T PHY design. Support is patchy to get access to most 10G PHYs (though this will improve with time).

Why not just use an SFP module, are you space limited?

Yes, I’m rather space limited (space on the PCB is limited, and I don’t have any space on the sides of the PCB, so if I add an SFP module, it not only takes the space for the cage, but also a big empty space to insert the module). In addition, as far as our experience goes, ethernet “cabling” is one of the least reliable parts of our robot (weather RJ45 connectors comming loose, or optical fiber getting damaged), so getting as may connections as possible onto our PCB is part of the reason we are designing this new PCB (and why we choose the Ubiswitch module, so we can get the signals directly onto our PCB instead of getting the RJ45 connections of a classical switch).

I see, this is indeed a bit of a pickle.

As it happens SFI and XFI are not directly compatible I believe. I couldn’t get a full confirmation on this so I went into the specification documents.

https://www.gigalight.com/downloads/standards/INF-8077i.pdf

https://www.gigalight.com/downloads/standards/sff-8431.pdf

My general conclusion is that they are “similar but different”; slightly different electrical specifications, so I don’t think XFI and SFI can be connected directly. Then I found the forum posts below.

https://e2e.ti.com/support/interface-group/interface/f/interface-forum/1091408/ds100df410-xfi-to-sfi-retimer

https://community.nxp.com/t5/Layerscape/LS1046-XFI-10G-design/m-p/1653022#M12514

Apparently they both use the same underlying encoding, but different electrical specification. Both posts suggest using a single retimer to convert between XFI and SFI. If that could be achieved, then this could work to make the connection between UbiSwitch and the Jetson

Of course, Nvidia come straight in and throw a bit of cold water on this idea. Not a straight “no” but a “we’re not sure” answer. I suppose they’re just covering themselves.

https://forums.developer.nvidia.com/t/ethernet-xfi-to-sfp/241342

Anyway from a design perspective, I guess you have the following options.

  1. Convert USXGMII from UbiSwitch to 10GBASE-T on your PCB using a 10GBASE-T PHY. Convert XFI from Jetson to 10GBASE-T using a 10GBASE-T PHY. Then connect the two 10GBASE-T links together on your PCB.

  2. Convert between USXGMII (which is the same as SFI) to XFI using a retimer as in those forum posts.

  3. Convert sgmii from UbiSwitch to 1GBASE-T on your PCB using a 1GBASE-T PHY. Convert XFI from Jetson to 1GBASE-T using a 1GBASE-T PHY. Then connect the two 1GBASE-T links together on your PCB.

  4. Use the RGMII port on Jetson, convert to 1GBASE-T, then connect that to UbiSwitch.

Honestly, option 4 is the least technical risk. Option 2 would be a good thing to try. From an engineering perspective, if this was my product, I’d design a product based on option 4 while in parallel, trying option 2. Of course, this is up to you.

We could also explore BotBlox doing a custom project on this for you, there’d be some NRE charge for that. We can explore that on a call if you would prefer that.