DRAM virtualizer MemVerge has teamed up with SK hynix on what they’re calling Project Endless Memory to run applications needing more memory than is actually attached to the server by using external memory resources accessed by CXL.
This is the Computer Express Link, the extension of the PCIe bus outside a server’s chassis, based on the PCIe 5.0 standard. CXL v2.0 provides memory pooling with cache coherency between a CXL memory device and several accessing server or other host systems. SK hynix has built CXL-connected DRAM products and MemVerge provides a software bridge to use them in the so-called Endless Memory project.
Charles Fan, CEO and co-founder of MemVerge, said in a statement: “The ability to download RAM automatically and intelligently will lead to improved productivity for users of data-intensive applications.”
Endless Memory combines MemVerge elastic memory service software and CXL-based Niagara Pooled Memory System hardware from SK hynix enabling host servers to dynamically allocate memory as needed to avoid running out of DRAM. Basically, if an app needs more DRAM than is physically present, the server can use MemVerge’s software to request more to be made available across CXL.
Hoshik Kim, VP/Fellow for Memory Forest R&D at SK hynix, said: ”Testing shows that just 20 percent extra CXL memory from our Niagara Pooled Memory System can improve application performance by 3x compared with the existing swap memory approach.” The swap memory approach is basically paging between physical memory and some kind of storage.
Memory Forest is an SK hynix concept. “Mem4EST” means memory for environment, society, and technology. Jonghoon Oh, EVP and CMO at SK hynix, described it like this: “Memory Forest represents how different layers of memory form a computing architecture, just like a forest, where everything from the bottom like the ground, grass to dense trees, fulfill their respective roles to form a huge ecosystem. Forests are generally, from bottom to top, divided into forest floor, understory, canopy layer, and emergent layer. These layers are naturally comparable to those of the computing architecture such as SSD (Solid State Drive), SCM (Storage-Class Memory), high-capacity memory (Capacity Memory), main memory and IPM (In Package Memory).”
Basically, SK hynix’s CXL expander memory hardware is included in this MemVerge initiative.
Gismo stands for Global IO-free Shared Memory Objects and is a CXL-based multi-server shared memory architecture for data access and collaboration in distributed environments. It enables real-time data sharing, using memory, across multiple servers, eliminating the need for network IO and so reducing data transfer delays.
Fan said: “Network and storage IO are performance bottlenecks for distributed, data-intensive applications. Project Gismo leverages the emerging CXL platform and will effectively remove this bottleneck known as the IO Wall. We envision a memory-centric future where memory is the network and memory is the storage.”
He suggests example use cases could be AI/ML applications, next-generation databases, and financial trading platforms.
One of the early adopters of Project Gismo is Timeplus, which is developing a real-time streaming database. Timeplus CEO Ting Wang said: “With Gismo’s revolutionary CXL-based multi-server shared memory architecture, we have experienced a remarkable improvement in the fault-tolerance of our database system. The speed of streaming query fail-over has been accelerated by an impressive 20x, allowing us to provide unparalleled user experience for the continuation of data streaming processing.”
MemVerge did not provide a timescale for Gismo completion.
Demonstrations of Endless Memory are being presented at the International Supercomputing Conference (ISC) in MemVerge’s booth, D401.