Broadcom Launches the Industry’s First 800G AI Ethernet NIC: Thor Ultra, Fully Compliant with UEC Standards

In a groundbreaking move for AI networking, Broadcom has unveiled the Thor Ultra, the industry’s first 800G AI Ethernet network interface card (NIC) chip that fully complies with the Ultra Ethernet Consortium (UEC) specifications. This innovative PCIe card enables seamless interconnections among hundreds of thousands of diverse XPUs, powering next-generation AI infrastructures with unprecedented speed and efficiency.

As of now, the only other 800G-speed RDMA NICs on the market are Broadcom’s Thor Ultra and NVIDIA’s CX-8 SuperNIC. Breaking through the limitations of traditional RDMA (RoCE) NICs, Thor Ultra introduces a suite of advanced features compatible with UEC standards, including packet-level multipathing, out-of-order packet delivery, hardware-based selective retransmission, telemetry, and fully programmable congestion control. Previously, AMD released the UEC-compliant 400G Pollara AI NIC, which outperformed InfiniBand in certain metrics. The 800G Thor Ultra NIC integrates seamlessly with Broadcom’s Tomahawk 6 switch chip and is compatible with any UEC-compliant switch products, making it ideal for Ethernet-based AI infrastructures.

Below, we’ve compiled key details from the Thor Ultra product manual to highlight its capabilities.

Key Advantages of Broadcom Thor Ultra 800G AI Ethernet NIC

Leveraging advanced RoCE features, the Thor Ultra significantly reduces completion times for AI workloads. Here’s how it delivers superior performance:

• Packet-Level Multipathing: By selecting the least congested path for each packet, it achieves efficient load balancing and maximizes network utilization.
• Out-of-Order Packet Delivery: Packets are delivered directly to XPU memory in any order, optimizing both network and PCIe bandwidth.
• Selective Retransmission: Hardware-accelerated selective retransmission ensures efficient data transfer with minimal overhead.
• Programmable Congestion Control: Users can choose receiver-based or sender-based congestion control mechanisms for flexible optimization.
• Advanced Telemetry: Works in tandem with Tomahawk 6 or any UEC-standard switch to enable packet trimming and congestion signaling.
• Effective Throughput of 800Gb/s: Equipped with PCIe Gen6 x16 lanes, it’s tailored for modern AI server architectures.
• Industry-Leading 100G SerDes: Extends direct-attach copper (DAC) cable reach up to 5 meters and supports Linear Pluggable Optics (LPO) to reduce optical module power consumption.
• Embedded Encryption Support: Offloads CPU-intensive encryption tasks, boosting overall system and application performance.
• Industry-Standard Management Interfaces and Protocols: Ensures easy integration and management in diverse environments.

Detailed Product Specifications

1) Network Interfaces

• Single-Port Speeds: Supports 800/400/200/100/50G rates.
• 8 SerDes Lanes: Compatible with 100/50G PAM4 and 25G NRZ signaling. According to Broadcom’s presentation at the OCP 2025 Summit, it supports configurations like 200G x4 lanes and 100G x8 lanes.
• Auto-Negotiation and Detection: Built-in features for seamless compatibility.

2) Host Interfaces

• PCIe Gen 6.0 x16 Lanes: Delivers high-bandwidth connectivity.
• Link Rates: 64, 32, 16, 8, 5, 2.5 GT/s.
• Supported Lane Configurations: x16, x8, x4, x2, x1 for flexible deployment.

3) RoCEv2 Network Features

• Packet-level multipathing for dynamic path selection.
• Out-of-order packet delivery to enhance throughput.
• Selective retransmission for reliable, low-latency transfers.
• Receiver- and sender-based congestion control.
• Adaptive routing and dynamic load balancing.
• Advanced telemetry for real-time monitoring.
• Peer Memory Direct (PMD) for direct memory access.
• Automatic configuration for simplified setup.

4) Network Virtualization and Acceleration

• Multi-Queue Support: Includes Multi-Queue, NetQueue, and VMQ for scalable processing.
• SR-IOV Virtualization: Enables efficient resource sharing.
• Virtual Function (VF) Isolation and Protection: Enhances security in virtualized environments.
• Tunnel-Aware Stateless Offload: Optimizes performance for encapsulated traffic.
• Stateless TCP Offload: Covers IP/TCP/UDP checksums, LSO (Large Send Offload), LRO (Large Receive Offload), GRO (Generic Receive Offload), TSS (TCP Segmentation Offload), RSS (Receive Side Scaling), aRFS (Accelerated Receive Flow Steering), and interrupt coalescing.

5) Application Scenarios

The Thor Ultra excels in demanding environments, including:

• Large-Scale AI/ML: Backend and frontend deployments for training and inference.
• High-Performance Computing (HPC): Accelerates compute-intensive workloads.
• Public Cloud Infrastructure: Scales efficiently for hyperscale providers.
• Private Cloud Infrastructure: Supports enterprise-grade deployments.
• Storage Systems: Enhances data transfer speeds in networked storage.

Broadcom’s UEC-First Approach vs. NVIDIA’s Spectrum-X Strategy

Unlike NVIDIA’s Spectrum-X Ethernet strategy, which ties into a proprietary ecosystem, Broadcom champions a “UEC-first” philosophy emphasizing openness. This approach provides crucial flexibility for operators wary of vendor lock-in, ensuring compatibility with UEC switches from multiple top-tier suppliers. As a result, the Thor Ultra NIC integrates effortlessly with diverse UEC-compliant ecosystems, promoting interoperability in AI Ethernet networking.

For those building Ethernet-based AI infrastructures, the Broadcom Thor Ultra represents a pivotal advancement in 800G AI Ethernet NIC technology. Its UEC compliance, combined with cutting-edge RoCE enhancements, positions it as a game-changer for reducing latency, boosting throughput, and scaling AI workloads. Stay tuned for more updates on Broadcom Thor Ultra, 800G AI networking solutions, and UEC-compliant Ethernet innovations. If you’re exploring options for your data center or cloud setup, Broadcom’s Thor Ultra could be the key to unlocking peak performance.