Compute Express Link, or CXL, has only been in use for five years, yet it is already having an impact in connecting server components.
The technology, introduced by Intel Corp. in 2019 and designed as an open-standard interface for high-speed communications, connects a server’s central processing unit with accelerator chips attached to the machine. As AI workloads increasingly place strains on server memory capacities, CXL has found appeal among firms seeking a performance boost.
Samsung’s David McIntyre and VMware’s Arvind Jagannath talk with theCUBE about CXL solutions.
“CXL itself really solves a problem on server constraints,” said David McIntyre (pictured, left), director of product planning and business enablement — device solutions — at Samsung Semiconductor, a subsidiary of Samsung Electronics Co. Ltd. “Servers can handle up to maybe five terabytes of dynamic random access memory, which is a huge amount of memory, but then it stops there. CXL provides memory expansion, eventually memory pooling. It basically frees up these constraints so that now you can expand out with additional devices or even systems that are CXL enabled. It’s living and thriving and it’s starting to find its way into actual production deployments.”
McIntyre spoke with theCUBE Research’s Dave Vellante and Rob Strechay at VMware Explore, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. He was joined by Arvind Jagannath (right), senior product line manager at VMware by Broadcom Inc., and they discussed how work by their companies is streamlining server performance.
Leveraging CXL solutions to avoid performance loss
VMware has been collaborating with Samsung researchers to deliver new innovations in memory. It participated with the company in March, when Samsung unveiled an expansion of its CXL memory module portfolio at a conference in Silicon Valley.
“We realized the potential of CXL very early,” Jagannath said. “We have a lot of expertise with memory in general because we were one of the leads in virtual machine migration, tracking pages and memory, et cetera. CXL has a lot of interesting features such as giving the CPU the ability to run instructions on the device directly. In terms of applications, we don’t want them seeing any performance loss, so CXL definitely achieves that for us.”
Samsung has been actively building modules to support CXL. The company introduced CMM-H, or CXL Memory Module-Hybrid. and also launched CMM-B, a memory pooling appliance for rack computing.
“This is putting a mix of the same types of devices all in one box that can be managed with Cognos, our orchestration layer,” McIntyre explained. “Samsung has now built a full portfolio of CXL-enabled platforms.”
VMware is looking at how memory hierarchy can support different AI workloads. Earlier this year, the company announced that vSphere 8.0 Update 3 would include support for Intel Xeon Max Series processors with integrated High Bandwidth Memory or HBM.
“We are looking into HBM as a potential tiering candidate because we have the expertise in tiering between different types of devices,” Jagannath said. “HBM could become a candidate between, for example, DRAM and HBM. That way we can make sure that a GPU gets additional memory because it’s tiering between HBM and DRAM.”
Here’s the complete video interview, part of SiliconANGLE’s and theCUBE Research’s coverage of VMware Explore: