Numem introduces AI memory engine with SRAM-class performance and 30–50% power savings

Numem has launched its AI Memory Engine, a memory subsystem IP designed to address the limitations of traditional SRAM and DRAM in AI and high-performance computing (HPC) workloads. The company identifies memory as a critical bottleneck, noting that while processing performance has increased 60,000X over the past 20 years, DRAM bandwidth has only improved 100X, creating a "memory wall" that restricts processor efficiency.

The AI Memory Engine is a fully synthesizable and configurable solution compatible with Numem’s patented MRAM-based architecture and third-party MRAM, RRAM, PCRAM, and Flash Memory. It delivers SRAM-class performance, up to 2.5X higher memory density in embedded applications, and 100X lower standby power consumption. The technology supports die density up to 1GB and is optimized for foundry-ready production, enabling integration into data centers and edge environments.
Key features include a flexible power management architecture with multiple power modes, high endurance for SRAM- and DRAM-like architectures, and scalable, software-defined memory that does not require hardware overhauls. The engine provides 30–50% power savings compared to existing high-bandwidth memory solutions, reducing operating costs ...