IN Brief:
- Semidynamics and Nokia have signed an exploratory memorandum, not a supply or deployment contract.
- Proposed work includes testing AI accelerator racks against Nokia Data Center Fabric networking.
- Semidynamics' RISC-V processors and Gazzillion memory subsystem are distinct from Nokia's network switching products.
Semidynamics and Nokia have signed a memorandum of understanding to examine how European AI accelerator hardware could operate alongside data centre networking systems. The proposed work covers assessment, testing and possible integration of Semidynamics’ liquid-cooled inference accelerator racks and Nokia’s Data Center Fabric technology.
Nokia’s Ethernet switching and management portfolio could provide the network interconnect for Semidynamics’ planned inference racks. From Barcelona, Semidynamics is developing processor and accelerator systems intended to run AI models, while Nokia supplies the fabric connecting servers across a data centre. The companies will evaluate how those layers work together before deciding on any integrated design or deployment.
Within the accelerator, Semidynamics combines a 64-bit RISC-V core with out-of-order execution, vector and tensor processing, and its Gazzillion memory subsystem. The architecture is intended to feed compute units with data while avoiding unnecessary movement between memory and processing elements. That internal data flow is separate from the Ethernet connections needed once multiple accelerator systems operate in one rack or across a larger installation.
The roadmap also includes a 3 nm inference SoC with multiple compute engines, a board linking host processors with inference devices through a high-bandwidth fabric, and persistent key–value cache storage for workloads with long context windows. These planned elements determine how model state and intermediate results may be held close to the processors. Nokia’s evaluation addresses communication beyond those internal compute and memory interfaces.
As a model generates successive tokens, previously calculated attention keys and values can be retained and reused. Longer contexts and larger numbers of simultaneous requests increase the state held in memory, making capacity and bandwidth as important as raw arithmetic throughput. A rack may therefore be limited by data movement even when its processors have unused compute capability. Distributing requests across multiple accelerators can add further transfers between devices, making the network topology part of the same performance calculation. The achievable balance depends on the workload, data layout, scheduling and the interconnect between processing elements.
Traffic leaving those processors would pass through a wider network architecture. Nokia’s Data Center Fabric brings together Ethernet switches and routers with SR Linux and Event-Driven Automation for configuration and operational monitoring. Leaf and spine switching connects groups of servers, providing paths for data exchange among compute nodes and onward to the rest of the data centre.
A network carrying concurrent inference requests must handle changing traffic volumes, latency and congestion without unnecessarily stalling the compute nodes. Performance will depend on where model data and results are held and how often traffic crosses the rack boundary. The planned evaluation will examine those interactions between Semidynamics hardware and Nokia’s fabric, with combined throughput and power figures depending on the eventual configuration and tests.
Power and thermal integration impose further design conditions. Semidynamics’ proposed rack design with liquid cooling is intended to follow Open Compute Project practices and must accommodate cooling distribution, electrical supply, cabling and system management alongside processor boards and network equipment. These interfaces will determine which configurations can be assembled and serviced if the two companies proceed beyond interoperability testing.
Separately, Semidynamics is collaborating with European processor company SiPearl on a rack-scale inference platform using SiPearl’s Arm host CPU and Semidynamics’ RISC-V acceleration technology. That project supplies an additional route for developing European compute hardware, while the Nokia memorandum concerns the network architecture with which such systems could communicate.
Assessment and validation of the proposed rack interconnect will determine how Semidynamics’ inference hardware can operate with Nokia’s existing networking products. Both companies will evaluate rack configuration, interconnect behaviour and operating characteristics before considering a commercial system. The separate SiPearl collaboration continues to address host computing and acceleration, while the Nokia evaluation concerns data centre connectivity.
