RACK-SCALE MEMORY DISAGGREGATION
Shared memory infrastructure for GPU clusters, high-performance computing, and data-intensive systems that outgrow fixed per-server memory.
Compare shared-memory options OUR
CHALLENGE
// PROBLEM: MEMORY BOTTLENECK
// TARGET: RACK-SCALE LATENCY
THE PROBLEM: MEMORY BOTTLENECK
Traditional servers tightly couple memory to individual CPUs and GPUs, creating bottlenecks and stranded capacity.
THE SOLUTION: RACK-SCALE MEMORY
TORmem removes this constraint by delivering a rack-scale Disaggregated Memory Fabric—a high-performance memory layer accessible by all servers in the rack.
Built on PCIe Gen5 and RDMA over RoCEv2 with ultra-low-latency switching, TORmem delivers near-local memory performance at rack scale—enabling large-memory workloads without per-node overprovisioning, proprietary hardware, or vendor lock-in.
The result: supercomputer-class memory scalability delivered through open, standards-based data center infrastructure.
MEMORY-CENTRIC
ARCHITECTURE
SUPPLY CHAIN
Streamlined supply chain management that eliminates concerns around component shortages and complex server memory configurations.
FUTURE PROOF
All servers can use the latest memory technologies, even over time.
EFFICIENCY
Decommissioned servers won't take large amounts of memory with them.
MEMORY-AS-A-SERVICE
Dynamic allocation enables true MaaS, allowing instant provisioning for AI machine leasing.
SOLVING
REAL PROBLEMS
MOORE'S LAW
The traditional approach to increase CPU compute capacity and local memory no longer scales. The amount of memory required for intensive use cases is growing, but space and cost are not.
MEMORY
UNLEASHED
IMC holds all required data within high-speed memory, eliminating the bottleneck of fetching data from slow storage. This keeps CPUs and GPUs fully fed.
SYSTEM EVOLUTION
LEGACY ARCHITECTURE
STATUS: INEFFICIENTHard choice between small amounts of fast memory and large amounts of slow storage.
TORMEM ARCHITECTURE
STATUS: OPTIMALNo longer limited by capacity. New workloads supported efficiently.
OPTIMIZING SYSTEM COSTS
By removing the need to populate each server with large amounts of local DRAM, TORmem reduces per-server system cost across the rack.
Through a rack-scale Disaggregated Memory Fabric, memory capacity is deployed where it is needed and shared efficiently, improving utilization while simplifying system design for both small and large deployments.
Patents
The patents behind the architecture
TORmem's name comes from the idea at the centre of its first patents: a top-of-rack memory server that pools memory and allocates it on demand to the other servers in the rack.
- US 11,507,285Granted Nov 22, 2022
Allocates memory from a shared pool and presents it to a server as an NVMe device, across different interconnect fabrics.
Systems and methods for providing high-performance access to shared computer memory via different interconnect fabrics
- US 11,561,697Granted Jan 24, 2023
A rack-mounted memory server that pools memory and allocates it on demand to other servers in the rack.
Disaggregated memory server having chassis with a plurality of receptacles accessible configured to convey data with PCIe bus and plurality of memory banks
- US 12,135,884Granted Nov 5, 2024
Continuation of the disaggregated memory server patent: rack-level memory pooling and on-demand allocation.
Disaggregated memory server having chassis with a plurality of receptacles accessible configured to convey data with PCIe bus and plurality of memory banks
- US 12,238,866Granted Feb 25, 2025
A high-density, liquid-cooled memory system: many memory modules and a processor on a half-width 1U board.
Systems and methods providing high-density memory arrangements with high-speed interconnects in a condensed form factor
Put the technology to work
See the memory-centric architecture in shipping hardware, or read how memory became the binding constraint on AI inference.
