Seamless integration with Micron CZ122 memory expansion modules and famfs delivers consistent, scalable performance for high-density AI and database applications

TAIPEI, May 20, 2025 – As cloud environments push infrastructure to its limits, memory bottlenecks remain a critical barrier to sustaining performance at scale. To address this, H3 Platform has introduced a powerful CXL Memory Sharing Solution, enabling multiple servers to share high-speed memory with low latency and zero software modification—delivering breakthrough efficiency for demanding workloads such as AI inference, real-time analytics, and large-scale databases.

The solution integrates Micron CZ122 CXL® memory expansion modules and operates with famfs (fabric-attached memory file system), allowing applications like RocksDB to run directly on shared meHyperscale

In hyperscaler-class deployments, where hundreds of workloads run in parallel, servers or compute nodes with traditional DRAM architectures struggle to maintain throughput once memory capacity is exceeded. H3 Platform’s architecture overcomes this by providing a shared, scalable memory pool accessible across multiple compute nodes via CXL.

Figure 1: Micron-generated results demonstrate the throughput comparison of RocksDB on traditional DRAM+SSD vs. CXL + famfs under growing dataset sizes.

In benchmark testing using RocksDB, DRAM-only systems showed sharp performance drops as datasets exceeded local memory. With DRAM and CXL + famfs, systems sustained stable throughput and low P99 latency, even under heavy memory pressure. This sustained throughput behavior can be beneficial to a multi-tenant environment. When CXL is added to a server running multi-tenant workloads, the server performance is improved by 70%, with less than 6% performance variance across VMs. This increased performance stems from the ability to dispatch more threads that can use the expanded memory when DRAM capacity is fully utilized. These results prove the solution’s ability to process dynamic, high-volume workloads consistently—a key requirement for modern AI and cloud services.

Architecture Overview: Designed for Composability and Speed

Figure 2: CXL memory architecture where multiple hosts access shared Micron CZ122 memory via FAMFS for RocksDB workloads.

The solution is designed with flexibility and ease of adoption in mind:

• Hardware Layer: DRAM combined with Micron CZ122 CXL memory expansion modules
• File System Layer: Famfs runs directly on shared memory, exposing it as files
• Application Layer: RocksDB and other tools operate directly atop famfs with zero code changes, sharing cacheline-granularity access to the memory

By enabling real-time, high-throughput access to shared memory across compute nodes, this architecture is optimized for hyperscaler-class AI inference, real-time analytics, and memory-bound services.

“H3 Platform’s mission is to build infrastructure that scales with the real needs of today’s data-intensive workloads,” said Brian Pan, CEO, H3 Platform. “With our CXL Memory Sharing Solution, we deliver consistent performance even in hyperscale scenarios—without asking customers to change the way they build software.”

“H3 Platforms delivery of the industry’s first switch-based CXL memory sharing solution, combined with the Micron CZ122 and famfs, enables the scaling of AI and database applications by eliminating memory bottlenecks,” said Vijay Nain, senior director of memory innovation and ecosystems at Micron Technology. “This collaboration allows our customers to achieve unprecedented efficiency and scalability, facilitating faster data processing, real-time analytics, and enhanced performance for their most demanding workloads.”

About H3 Platform

H3 Platform is a technology leader in composable infrastructure solutions based on PCIe and CXL. By developing scalable, memory-centric architectures, H3 enables enterprises to efficiently power AI, HPC, and cloud workloads across modern data centers. The company’s CXL-based memory sharing technologies simplify resource orchestration, reduce total cost of ownership, and deliver consistent performance for next-generation applications.

Learn more at www.h3platform.com