AMD will be the first to market with a new Ultra Ethernet-based networking card, and Oracle will be the first cloud service provider to deploy it.
The announcement came at the recent Advancing AI event, where AMD introduced its latest Instinct MI350 series GPUs and announced the MI400X, which will be delivered next year. Overlooked in that news blitz is the availability of the Pensando Pollara 400GbE network interface card, which marks the industry’s first NIC that’s compliant with the Ultra Ethernet Consortium’s (UEC) 1.0 specification.
AMD announced Pollara in 2024, but it is only just beginning to ship it. And just as the 400Gb Pollara begins shipping, AMD also announced a next generation 800Gb card dubbed Vulcano, which is also UEC-compliant. AMD’s announcement came just days after the UEC published its 1.0 specification for Ultra Ethernet technology, designed for hyper-scale AI and HPC data centers.
The UEC was launched in 2023 under the Linux Foundation. Members include major tech-industry players such as AMD, Intel, Broadcom, Arista, Cisco, Google, Microsoft, Meta, Nvidia, and HPE. The specification includes GPU and accelerator interconnects as well as support for data center fabrics and scalable AI clusters.
AMD’s Pensando Pollara 400GbE NICs are designed for massive scale-out environments containing thousands of AI processors. Pollara is based on customizable hardware that supports using a fully programmable Remote Direct Memory Access (RDMA) transport and hardware-based congestion control.
Pollara supports GPU-to-GPU communication with intelligent routing technologies to reduce latency, making it very similar to Nvidia’s NVLink c2c. In addition to being UEC-ready, Pollara 400 offers RoCEv2 compatibility and interoperability with other NICs.
At the Advancing AI event, AMD CEO Lisa Su introduced the company’s next-generation, scale-out AI NIC, Vulcano. Vulcano is fully UEC 1.0 compliant. It supports PCIe and dual interfaces to connect directly both CPUs and GPUs, and it delivers 800 Gb/s of line rate throughput to scale for the largest systems.
When combined with Helios – AMD’s new custom AI rack design – every GPU in the rack is connected through the high-speed, low-latency UA link, tunneled over standard Ethernet. The result is a custom AI system very similar to Nvidia’s NVL-72, where 72 GPUs are made to look like a single processor to the system.
Oracle is the first to line up behind Pollara and Helios, and it likely won’t be the last. Oracle lags the cloud leaders AWS and Microsoft and only has about 3% of the public cloud market.
Source:: Network World