Arrcus and UfiSpace Partner on AI Networking
Arrcus and UfiSpace have announced a collaboration on AI-optimized networking platforms that span from the edge to the data center core. The partnership highlights the growing importance of the network stack as a critical performance bottleneck in large-scale AI clusters.
The partnership between Arrcus and UfiSpace centers on a disaggregated networking model, separating hardware and software to give customers more flexibility. This approach counters the traditional, vertically integrated systems from single vendors. The collaboration will feature Arrcus's ACE-AI networking software running on UfiSpace's open hardware, designed for AI workloads from the network edge to the data center core. At the heart of the joint offering are UfiSpace's S9700 and S9710 series switches, which utilize Broadcom's Jericho2c+ and Jericho3-AI silicon. The Jericho3-AI chip, for example, can support network clusters of up to 32,000 GPUs at 800Gbps speeds. Arrcus's ArcOS, in turn, provides the unified software layer across these different hardware platforms. This collaboration directly addresses the unique traffic patterns of AI clusters, which differ significantly from traditional data center workloads. AI applications often create "incast congestion," where many GPUs attempt to send data to a single point simultaneously, overwhelming network switches. Technologies like scheduled fabric and advanced load balancing in the Jericho3-AI are designed to prevent this congestion and reduce job completion times. The competitive landscape for AI networking is intense, with major players like Nvidia and Arista making significant moves. Nvidia's Spectrum-X platform offers a full-stack solution combining switches, NICs, and software, which it claims improves AI network performance by 1.6 times over traditional Ethernet. Arista, in contrast, partners with Broadcom for its silicon and focuses on its EOS software to manage AI workloads. This makes the Arrcus and UfiSpace partnership a key example of the open, disaggregated ecosystem model competing against more proprietary solutions.
