AI is driving a fundamental rethink of data center operations. Real-time training and inference demand ultra-low latency, while even minor packet loss can derail training jobs and delay completion. Combined with massive bandwidth requirements, these pressures are pushing existing data center networks to their limits.
Ultra Ethernet Consortium (UEC) Specification 1.0 defines this new UET layer to solve the complex problem set that HPC/AI creates. With Ethernet widely adopted across most data networks worldwide, it is a compelling protocol for modernizing AI data center networks with new standards, best practices, and architectures to support AI and HPC workloads. As an active member of the UEC, Nokia is committed to leveraging UET as a highly compatible, cost-effective, and interoperable part of AI and HPC application stacks.
he procedure tested 800 GE interfaces with flows of UET traffic generated by Keysight’s AresONE 800GE-8P-QDD-M testing platform over a network that spanned all variants of the Nokia 7220 IXR-H5 and the Nokia 7250 IXR-10e platforms. These Nokia switches were running the SR Linux Network Operating System (NOS).
As Nokia integrates UEC Specification 1.0 capabilities into the Nokia data center fabric solution for HPC/AI networks, global customers already have immediate commercial access to an AI-ready and UEC-compatible portfolio with support for Remote Direct Memory Access over Converged Ethernet (RoCEv2) and Data Center Quantized Congestion Notification (DCQCN). The current support for RoCEv2 was also tested at 800 GE flows in parallel on the same testbed, showcasing coexistence, openness, and flexibility.
