PCIe Gen6 and CXL 3.0/ CXL 3.1 are happening, and soon. As a result, we have new demos with products like the new switch chips from XConn Technologies. This is a startup taking on industry giants like Broadcom for next-gen PCIe and CXL infrastructure, so its chips are neat. We previously saw the company’s chips at events like last year’s FMS 2024.
XConn Tech Shows off New PCIe Gen6 and CXL 3 Switch Chips at FMS 2025
On the show floor, XConn Tech had a demo running, which was very neat. Here we can see a switch development board with a host system and a CXL memory device.
This is XConn Tech’s next-gen Apollo2 design. We can see that this switch chip is a PCIe Gen6 enabled switch.
The other thing is that we can see that this is enumerating the ScaleFlux CXL 3.0 device even over PCIe Gen5. We think this is actually a CXL 3.1 device, but it was showing 3.0 on this demo.
In the PCIe Gen5 era, we focused on the XConn 256-port switch. Now, the company is building a larger portfolio of switch ports like the XC60064. For reference, the XConn XC50256 was a Gen5 256-port switch. Following that naming, this is a Gen6 64-port switch.
XConn is talking about the ability to give GPU direct access to a CXL memory pool using a hybrid CXL/PCIe hybrid switch.
Folks in the industry are grappling with the fact that modern AI is consuming a lot of high speed memory. The cost of using HBM3E/ HBM4 is so high that anything that can move to lower speed tiers of memory for capacity is useful.
XConn was also talking about the price comparison for a CXL memory pool.
The big dollar change is when we get memory shelves that can be partitioned dynamically in the future.
Final Words
It is always neat getting to see next-generation parts. PCIe Gen6 is rapidly approaching as we have already discussed NVIDIA B300 is PCIe Gen6. Now we need next-generation servers to get there so we can use cool parts like this.
The Direct Connect CXL is excellent, super large AI models with trillions of parameters:
https://www.linkedin.com/posts/pacific-northwest-national-laboratory_check-out-this-new-active-memory-computing-ugcPost-7364340617369317381-9op-?utm_source=share&utm_medium=member_desktop&rcm=ACoAAET3EtUBlHV4p8XM378EshTbSOQ1CbUCnV0
https://www.pnnl.gov/news-media/unique-active-memory-computer-purpose-built-ai-science-applications
Faster, cheaper and more efficient was the theme in computer engineering for the last 80 years until AI made faster, more expensive and power hungry popular.
Around here a cluster of a 100 Haswell-generation Xeon servers with P100 accelerators continues to do science and engineering as if caught in a time warp. Although the storage was upgraded this summer, the rest of the machine just gets more and more reliable over time.
Since PCIe6 will be out soon, maybe it’s lucky those budget constrains caused PCIe4 and PCIe5 to be skipped!