Broadcom Inc. introduces Thor Extremely, claiming the trade’s first 800G AI Ethernet community interface card (NIC). The Ethernet NIC, adopting the open Extremely Ethernet Consortium (UEC) specification, can interconnect tons of of hundreds of XPUs to drive trillion-parameter AI workloads.
The UEC modernized distant direct reminiscence entry (RDMA) for giant AI clusters, which the Thor Extremely leverages, providing a number of RDMA improvements. These embrace packet-level multipathing for environment friendly load balancing, out-of-order packet supply on to XPU reminiscence for maximizing cloth utilization, selective retransmission for environment friendly information switch, and programmable receiver-based and sender-based congestion management algorithms.
By offering these superior RDMA capabilities in an open ecosystem, it permits prospects to connect with XPUs, optics, or switches and to scale back dependency on proprietary, vertically built-in options, Broadcom stated.
The Thor Extremely joins Broadcom’s Ethernet AI networking portfolio, together with Tomahawk 6, Tomahawk 6-Davisson, Tomahawk Extremely, Jericho 4, and Scale-Up Ethernet (SUE), as a part of an open ecosystem for giant scale high-performance XPU deployments.
The Thor Extremely Ethernet NIC is obtainable in commonplace PCIe CEM and OCP 3.0 type components. It presents 200G or 100G PAM4 SerDes with assist for long-reach passive copper, and claims the trade’s lowest bit error charge SerDes, decreasing hyperlink flaps and accelerating job completion time.
Different options embrace a PCI Specific Gen 6 ×16 host interface, programmable congestion management pipeline, safe boot with signed firmware and system attestation, and line-rate encryption and decryption with PSP offload, which relieves the host/XPU of compute-intensive duties.
The Ethernet NIC additionally gives packet trimming and congestion signaling assist with Tomahawk 5, Tomahawk 6, or any UEC compliant swap. Thor Extremely is now sampling.
The submit AI Ethernet NIC drives trillion-parameter AI workloads appeared first on EDN.
