A Chinese vendor showed direct chip-to-network-card links on all-domestic hardware

Kiwimoore and Biren demonstrated a direct chip-to-network-card link in Shanghai last week, running on all-domestic silicon rather than the Nvidia ConnectX NICs that have handled the job in every other IBGDA setup disclosed so far. The work, shown at the World AI Conference 2026 (July 17–20), is the first publicly described instance of a Chinese AI accelerator talking to a Chinese RDMA NIC in IBGDA mode without an Nvidia part in the path.

The demonstration matters because of which layer it advances. Modern AI training depends on the network that wires accelerators together, not the chips themselves, but the cards and switches between them. RDMA, or Remote Direct Memory Access, lets one machine write into another's memory without involving the operating system, and the network card is the piece that holds that door open. IBGDA, or InfiniBand GPU Direct Async, is the protocol that lets a GPU launch a network operation directly to a network card, skipping the CPU bounce that bottlenecks modern MoE training, where Mixture-of-Experts models constantly shuffle which experts are doing the work. Until now, domestic Chinese clusters that wanted IBGDA got it by buying Nvidia ConnectX cards and pairing them with Chinese GPUs. The Kiwimoore–Biren run replaces that pair with a Biren GPU talking to a Kiwimoore Kiwi SNIC 800G card, a 400G RDMA ASIC engine, on an open Ethernet stack.

The numbers, as Kiwimoore presented them on the WAIC show floor, are pointed at a specific workload: the All-to-All traffic that fires during MoE expert-parallel Dispatch and Combine, when every GPU in the cluster must send a fragment of every token to whichever expert handles it. Against a CPU-bounce IBRC baseline, the company says its small-packet single-message bandwidth "significantly" improved, raw and aggregated small-packet throughput crossed thresholds, and end-to-end All-to-All latency fell by nearly half. The "small-packet penalty" that haunts MoE collectives is, per the vendor, eliminated. The figures align with the public numbers Nvidia has published for ConnectX running the same workload in IBGDA mode, which is the relevant benchmark for a company that wants to claim parity at the layer Nvidia owns.

The technology stack is open Ethernet rather than InfiniBand. The company is leaning on the same general direction the hyperscalers have taken since 2023, building large training fabrics on commodity Ethernet plus RDMA rather than Nvidia's proprietary switch and adapter stack. Kiwimoore's pitch, delivered by 祝俊东 (Zhu Jundong) at the AI Network Frontier Ecosystem Forum, is that this gives domestic operators a "system-level infrastructure" that is "evolvable, collaborative, and unbundled." The same event saw a national super-node initiative launched and a co-packaged optics white paper released, with participation from Shanghai AI Lab, HKUST (Guangzhou), SenseTime, Biren, DataCanvas, Sunlune, Kehua Cloud, Turing Quantum, and Turing Chip.

What the demo does not yet show is parity with Nvidia's networking stack under production load. A trade-show run on a single workload is the smallest unit of evidence a serious claim requires. Nvidia's ConnectX has years of production tuning, a mature NCCL and NVSHMEM upper layer, and a deep install base. The Kiwimoore card has the silicon and the demo, plus compatibility with a DeepEP-style MoE collective library, but the published numbers do not yet include independent benchmarks, full training runs, or disclosed software versions and traffic mixes. The "small-packet penalty" the vendor says it has eliminated is one of the harder problems in MoE networking, and the community will want to see the result reproduced outside a Kiwimoore press event.

The strategic read is narrower than the headline suggests. The demo advances the connectivity layer of China's AI stack onto a domestic roadmap, which matters for planning and for procurement under export controls. It does not mean Chinese AI clusters no longer need Nvidia networking. It means there is now a domestic option for the inter-accelerator link that did not exist in IBGDA form a week ago. Whether that option scales to full training clusters is the next question, and the WAIC main stage is the wrong room to settle it. The first independent benchmarks, on disclosed code, are the ones that will.