AI Ethernet NIC drives trillion-parameter AI workloads

Broadcom Inc. introduces Thor Ultra, claiming the industry’s first 800G AI Ethernet network interface card (NIC). The Ethernet NIC, adopting the open Ultra Ethernet Consortium (UEC) specification, can interconnect hundreds of thousands of XPUs to drive trillion-parameter AI workloads.

The UEC modernized remote direct memory access (RDMA) for large AI clusters, which the Thor Ultra leverages, offering several RDMA innovations. These include packet-level multipathing for efficient load balancing, out-of-order packet delivery directly to XPU memory for maximizing fabric utilization, selective retransmission for efficient data transfer, and programmable receiver-based and sender-based congestion control algorithms.

By providing these advanced RDMA capabilities in an open ecosystem, it allows customers to connect to XPUs, optics, or switches and to reduce dependency on proprietary, vertically integrated solutions, Broadcom said.

(Source: Broadcom Inc.)

The Thor Ultra joins Broadcom’s Ethernet AI networking portfolio, including Tomahawk 6, Tomahawk 6-Davisson, Tomahawk Ultra, Jericho 4,  and Scale-Up Ethernet (SUE), as part of an open ecosystem for large scale high-performance XPU deployments.    

The Thor Ultra Ethernet NIC is available in standard PCIe CEM and OCP 3.0 form factors. It offers 200G or 100G PAM4 SerDes with support for long-reach passive copper, and claims the industry’s lowest bit error rate SerDes, reducing link flaps and accelerating job completion time.

Other features include a PCI Express Gen 6 ×16 host interface, programmable congestion control pipeline, secure boot with signed firmware and device attestation, and line-rate encryption and decryption with PSP offload, which relieves the host/XPU of compute-intensive tasks.

The Ethernet NIC also provides packet trimming and congestion signaling support with Tomahawk 5, Tomahawk 6, or any UEC compliant switch. Thor Ultra is now sampling.

The post AI Ethernet NIC drives trillion-parameter AI workloads appeared first on EDN.