This week, I stopped by the Xsight Labs office in California. While I was there, I was able to see the company Xsight Labs E1 DPU, but also the Xsight Labs X2 12.8T network switch. Recently, the company announced a new universal switch with Edgecore. While I did not get to see that version, having spent the last three months getting our high-end traffic generation machine working, I can understand the appeal.
Xsight Labs X2 12.8T Network Switch Goes Universal
The actual switch chip is much smaller than a higher-end chip like the Broadcom Tomahawk 6. The idea is not to replace the core switching at AI clusters. Instead, this is more of a top-of-rack (TOR) switch chip.
Here is the other side of the switch chip.
The Xsight Labs X2’s block diagram is below:
One of the advantages of the platform is that it can run in under 200W. It also supports the industry standard SONiC software making it easier to integrate. Part of what Xsight Labs is also doing is making a switch chip that can handle lower speed ports. Split out to 100GbE, it can support 128 connections which is substantial in that power envelope.
The switch you are seeing is the development platfom, but I thought folks might like seeing it since we did not get to see the Edgecore version.
The universal switch platform uses 128x 112G XLR (extra long range) SerDes, so it supports QSFP112. That may not seem like a big deal, but it means you can have four lanes to make 400GbE instead of eight if you used 56G SerDes. Practically, that means it is very easy to have dual 200G using two sets of two of these and using four split you can get four 100G ports. While that might seem trivial, we have been hunting down the right optics for our 400G/800G setups and that is not a small task. It is also one many run into these days.
Final Words
Hopefully we will get to check out the X2 in the future, and maybe this version. Still, it is important to see new switch silicon hit the market because we have seen some companies exit. It is a tough business when you have companies like Broadcom, NVIDIA, Cisco, Marvell, and others as competitors. Those are also companies very focused on the 51.2T and 102T switch generations for large-scale AI clusters at the moment, leaving room for innovation in other segments.
While I was there, a box arrived with the first of a new product that will likely be shown at OCP 2025. Stay tuned for that, as it was super cool.
Excited to put the X2 into the Oxide computer!
Was the new box maybe the dual E1 DPU + CPU in a 1U chassis version? That would be really cool!
BTW Why didn’t you mention the open ISA (machinecode) and because of that, the open programmability of the swich? Also something to write a separate article about!
That is exciting Kevin. Also SanderT – Hoping to do more hands-on with these. I saw some cool demos.
Why wouls anyone be interested in a 2018 technology?
This is so obsolete.