Nvidia-supported company develops Ethernet memory reservoir for AI enhancement, asserting it can offer up to 18TB of DDR5 storage for substantial inference jobs and potentially slash per-token creation expenses by half.
Enfabrica Unveils Game-Changing AI Memory Fabric System
In a groundbreaking move, Enfabrica has launched its Elastic Memory Fabric System, dubbed Emfasys, on July 29, 2025. This innovative system aims to revolutionise the way large-scale AI inference workloads are handled[1].
At the heart of Emfasys is a solution to the memory bandwidth-scaling and memory margin-stacking problems that AI inference currently faces, as stated by Rochan Sankar, CEO of Enfabrica[1]. The system adds terabytes of DDR5 memory to any server via an Ethernet connection, making it rack-compatible[1].
The Emfasys system is built upon the ACF-S SuperNIC, boasting a 3.2 Tb/s throughput, and connects up to 18 TB of DDR5 memory with CXL[1]. It allows data movement between GPU servers and the Emfasys memory pool using Remote Direct Memory Access (RDMA), ensuring zero-copy, low-latency memory access without CPU intervention[1].
Enfabrica's Emfasys is currently being evaluated and tested by select clients[1]. In high-turn and long-context scenarios, using the Emfasys memory pool can lower the cost per AI-generated token by as much as 50%[1].
The system is designed to address the growing memory requirements of modern AI use cases that employ increasingly long prompts, large context windows, or multiple agents[1]. To access the Emfasys memory pool, servers require memory-tiering software provided or enabled by Enfabrica[1].
Servers can access the memory pool through standard 400G or 800G Ethernet ports using RDMA over Ethernet[1]. By using the Emfasys memory pool, data center operators can flexibly expand the memory of an individual AI server[1].
Enfabrica is not only a member of the Ultra Ethernet Consortium (UEC) but also contributes to the Ultra Accelerator Link (UALink) Consortium[1]. This software runs on existing hardware and OS environments and builds upon widely adopted RDMA interfaces[1].
While the availability of the Emfasys system for general use is currently unknown, the launch marks the system's initial introduction to the market and indicates that further coverage or updates about Enfabrica and Emfasys may be forthcoming[1].
[1] Enfabrica Press Release, July 29, 2025.
Emfasys, the innovative memory fabric system developed by Enfabrica, relies on advanced technology to deliver a solution for the memory bandwidth-scaling and memory margin-stacking problems faced by AI inference workloads. This technology-driven system leverages data-and-cloud-computing techniques to revolutionize the handling of large-scale AI inference workloads.
