SOLUTION
EMFASYS:
Elastic AI Memory Fabric System
- Industry's first RDMA Ethernet based AI Memory Fabric architecture
- Offloads GPU, HBM, and local head node DDR resource consumption
- Drives down the cost of LLM inference at fleet scale
- Seamless integration with popular LLM inference applications frameworks
Under the Hood
Under the Hood
- Based on Enfabrica's 3.2 Tbps ACF SuperNIC silicon
- Elastically connects up to 144 CXL memory lanes per system to resilient bundles of 400G / 800G RDMA Ethernet ports
- Pooled memory target up to 18 TB CXL DDR5 DRAM per system (expandable to 28 TB in the future)
- High memory bandwidth aggregation enabled by striping transactions across 18+ memory channels per system
- Read access times in microseconds
- Designed to interoperate with multiple GPU servers and initiator RDMA NICs
System design for reduced cost per token per user
System design for reduced cost per token per user
- Ideal for agentic, batched, expert-parallel, high-turn, and/or large-context inference workloads
- Applies to AI training: activation storage offload, distributed checkpointing, optimizer state sharding
- Software-enabled, high-performance caching hierarchy hides transfer latency within compute pipelines
- Designed to outperform flash-based inference storage offload solutions with 100x lower latency and unlimited write/erase transactions
- 100% fabric-attached, headless memory free from head node CPU thread contention or locality constraints
