Broadcom BCM57508 200Gb Dual-Port Ethernet Adapter (Thor)

The Broadcom BCM57508 200Gb Dual-Port Ethernet Adapter, known by its Thor architecture codename, is a fifth-generation SmartNIC built on Broadcom's advanced 7nm ASIC process. Each of the two ports delivers 200 Gigabit Ethernet throughput, yielding a combined 400 Gbps of bidirectional bandwidth from a single PCIe Gen 4 x16 host slot. The adapter implements the full RoCEv2 (RDMA over Converged Ethernet) protocol stack in hardware, offloading kernel bypass communications entirely from host CPU cycles and enabling sub-microsecond latency for tightly coupled distributed workloads.

The BCM57508 integrates Broadcom's fifth-generation P5 core engine with hardware-accelerated stateless offloads including Large Send Offload (LSO), Large Receive Offload (LRO), RSS, TSS, and full checksum offload for TCP, UDP, and IP. The adapter supports SR-IOV with up to 256 virtual functions per port, enabling dense multi-tenant virtualization and container networking at scale. Line-rate flow classification, VXLAN and GENEVE tunnel offload, and precise hardware timestamping for IEEE 1588 PTP are implemented directly in the ASIC, reducing host CPU overhead across all traffic profiles.

Designed expressly for AI and machine learning infrastructure, the BCM57508 accelerates all-reduce collective operations across GPU clusters by pairing with NVIDIA GPUDirect RDMA, allowing GPU memory to be accessed directly over the fabric without staging through host memory. It is equally suited to high-frequency trading platforms, distributed storage back-ends such as NVMe-oF, and hyperscale cloud east-west traffic where consistent, low-latency throughput at scale is mandatory. The adapter ships in a standard PCIe low-profile form factor, integrating smoothly into 1U and 2U rack servers from major platforms.

Ideal for

• Distributed AI and machine learning training clusters requiring GPU-to-GPU RDMA collective communications (all-reduce) across 200GbE fabrics with GPUDirect RDMA support
• NVMe over Fabrics (NVMe-oF) storage back-end connectivity in all-flash arrays and disaggregated storage architectures demanding consistent sub-microsecond latency
• High-performance computing (HPC) job scheduling environments where MPI workloads benefit from hardware-offloaded RoCEv2 and lossless Ethernet via DCQCN congestion control
• High-frequency trading and capital-markets platforms leveraging IEEE 1588 PTP hardware timestamping and kernel-bypass RDMA for deterministic, ultra-low-latency order processing
• Dense multi-tenant cloud and virtualization infrastructure using SR-IOV with up to 256 virtual functions per port to provide isolated, line-rate networking to VMs and containers
• Hyperscale east-west data-center interconnect with VXLAN and GENEVE hardware tunnel offload, sustaining full 200GbE throughput across overlay networks without CPU involvement

Technical specifications

Available from Omnixon Global. Submit an RFQ and our team will confirm configuration and availability for your order.