CXL Protocol Adds Capabilities over PCIe

[erek]

Opinion? This seems cool.

"As an aside, he said these new synergies have prompted the reimaging of another SNIA endeavor, the Solid State Storage Initiative, in favor of something that supports persistent memory storage, computational storage, CXL and other interfaces.

CXL’s appeal and quick adoption has not only been driven by the fact that it runs over the existing PCIe layer. There’s also a commitment to the “KISS” principle, Shadley said, so it’s easy for developers to deploy solutions, as well as lots of collaboration between the groups representing the different yet relevant technologies such as NVMe and Gen-Z. “It’s the first open interface and protocol that’s being driven by everybody that’s of significance in the market from a compute perspective,” he said."

https://www.eetimes.com/cxl-protocol-adds-capabilities-over-pcie/

 

[Lakados]

It's based off the PCIe 5.0 specification so we likely won't see it for a while in the desktop arena but overall it is a cool project, I am Glad AMD joined in on it.
Here's a nicer link to a break down of what the CXL does.
https://www.synopsys.com/designware-ip/technical-bulletin/compute-express-link-standard-2019q3.html

 

[erek]

It's based off the PCIe 5.0 specification so we likely won't see it for a while in the desktop arena but overall it is a cool project, I am Glad AMD joined in on it.
Here's a nicer link to a break down of what the CXL does.
https://www.synopsys.com/designware-ip/technical-bulletin/compute-express-link-standard-2019q3.html

Thank you

 

[Thunderdolt]

It sounds interesting and I like the concept, but I don't think it has much applicability outside of the datacenter. When my system is running out of VRAM, it usually doesn't have tons of system RAM available either. If the Intel desktop chips weren't neutered and could handle more RAM, then this would become more useful. The promise is huge, it's just that there are currently other bottlenecks which will significantly reduce its utility until they're resolved.

If this allowed coherence between GPUs, I think I'd find it more interesting - but this bit is probably better solved by bridging GPUs ala NVLink. Because it's such a long way off, Nvidia's Ampere bridging technology is likely to be at least as fast and Hopper's will likely be considerably faster than that.