Nvidia moves AI to the edge

Nvidia’s latest push is moving high-end artificial intelligence computing out of data centres and into factory racks, mobile networks and other field hardware. (eurotech.com) Eurotech said on April 16 that its ReliaCOR 55-20 and 61-11 edge servers are now available for mass production with Nvidia Blackwell graphics processors and Advanced Micro Devices EPYC server chips. The company said the systems target visual inspection, medical imaging, traffic analysis, safety analytics and other workloads that need local inference instead of a remote cloud round trip. (eurotech.com) The hardware is built for tight spaces rather than warehouse-scale server halls. Eurotech said the chassis is under 470 millimeters deep, while the 2U ReliaCOR 55-20 supports one Nvidia RTX Pro 5000 or 6000 Max-Q Blackwell graphics processor and uses redundant power supplies, fans and removable disks for industrial deployments. (eurotech.com 1) (eurotech.com 2) A radio access network is the part of a mobile system that links phones to antennas and base stations. Nokia and Orange said on April 15 they are testing artificial intelligence radio access network, or AI-RAN, using Nokia anyRAN 5G software and Nvidia infrastructure to improve spectral efficiency, energy use and services such as integrated sensing and communication. (nokia.com) Nokia said the work will examine a graphics-processor-based radio processor, predictive optimization and artificial-intelligence-assisted functions such as scheduling, beamforming and power optimization. Orange said the program fits its network plans across Europe, the Middle East and Africa and is intended to support a software-defined path toward sixth-generation wireless systems. (nokia.com) The edge pitch is simple: run the model where the camera, machine or antenna already sits. That cuts latency, reduces backhaul traffic and keeps more data on site, which is why vendors are now trying to package data-centre-class accelerators into shorter, serviceable systems for hospitals, factories and city infrastructure. (eurotech.com) (nokia.com) Nvidia is also extending that same computing stack into quantum-adjacent work, where artificial intelligence is used to tune and stabilize fragile experimental processors. On April 14, Nvidia introduced its open-source Ising model family for quantum calibration and error correction, two bottlenecks that determine whether quantum hardware can run useful jobs reliably. (nvidianews.nvidia.com) Nvidia said Ising can cut calibration time from days to hours and improve error-correction decoding with up to 2.5 times faster performance and 3 times higher accuracy than traditional approaches. The company said users already include Harvard, Lawrence Berkeley National Laboratory, Fermi National Accelerator Laboratory, IQM Quantum Computers and the United Kingdom National Physical Laboratory. (nvidianews.nvidia.com) Taken together, the announcements show Nvidia selling the same basic idea in three places at once: inside industrial boxes, inside telecom networks and alongside quantum control systems. The common thread is that more artificial intelligence processing is being pushed closer to the machine that produces the data. (eurotech.com) (nokia.com) (nvidianews.nvidia.com)