Rambus (RMBS) Status Update summary

Market drivers and challenges

• AI and cloud workloads are rapidly increasing demand for both memory bandwidth and capacity, with larger and more sophisticated models requiring ever more resources.

• Modern AI training and inference pipelines require main memory and GPU memory capacities ranging from less than 1TB to up to 8TB.

• CPU core counts are rising dramatically, with new generations exceeding 100 cores and projections for over 200, intensifying memory subsystem bottlenecks.

• Cloud infrastructure faces a "memory wall" as memory bandwidth struggles to keep pace with rising CPU core counts.

• Physical constraints on server motherboards limit the ability to add more traditional DIMMs, necessitating new solutions for higher bandwidth and capacity.

MRDIMM technology overview

• MRDIMM leverages existing DDR5 infrastructure, using the same DRAM but adding data buffers to multiplex data streams, effectively doubling bandwidth per channel.

• MRDIMMs introduce 10 data buffers (MDBs) and a new Registering Clock Driver (MRCD) for enhanced performance.

• MRDIMM modules can start at 12,800 MT/s and scale higher, extending the DDR5 roadmap and supporting greater capacity through new form factors and packaging.

• The architecture allows DRAM data to be multiplexed to a 12.8GT/s host interface, doubling the native speed of DDR5 DRAMs.

• Compatibility is maintained, allowing MRDIMM and RDIMM modules to be used interchangeably in future systems.

Performance and infrastructure benefits

• MRDIMMs extend DDR5 data rates from 9200 MT/s up to 12800 MT/s, with future generations targeting 18000+ MT/s.

• They leverage existing DDR5 infrastructure, allowing seamless upgrades and extending DDR5's lifecycle.

• Rambus offers the industry's first MRCD and MDB chips for MRDIMMs, enabling flexible, scalable server configurations.

• MRDIMMs support both standard and tall DIMM form factors, enabling greater capacity per channel.

• The solution incorporates advanced clocking, control, and power management for high-capacity, high-bandwidth modules.