Shenzhen-based chip designer Lanxin Computing has launched what it calls a converged-computing supernode, extending its RISC-V AI inference platform from a single chip and server to a cluster-level system. The company presented the product at the 2026 RISC-V China Summit, where chief technology officer Jia Xinxia gave a keynote on the LX500 processor that underpins it.
The idea behind the design is to cut down on data movement. In a conventional inference server, a CPU and a separate accelerator card each have their own memory, and model weights, activations and context data shuttle between them over PCIe. Lanxin's approach integrates general-purpose computing and AI computing on the same chip and lets both share a unified memory pool, which the company says helps with time to first token, prefill and long-context workloads.
According to the company, LX500 packs 48 RISC-V CPU cores, split into 32 performance cores and 16 efficiency cores, alongside AI vector and matrix engines rated at 75 TOPS of INT8 compute. It supports precisions from INT4 to FP32, shares DDR5 memory across the chip, and supports CXL memory expansion. Earlier company materials say the chip follows the RVA23 profile and the RVV 1.0 vector extension.
The supernode caps a product line that Lanxin has assembled over this year. LX500 was powered on in February, shown publicly in April, and used in an all-in-one prototype in May. In July, Lenovo's WR5219 G5 server built on the chip was unveiled, and in August the company began sampling LX500 to industry customers, some of which have moved into small-batch deployment testing. The current lineup runs from the chip and server to the LX5000 all-in-one appliance and the new supernode.
Lanxin did not disclose how many chips a supernode holds or what interconnect links them, so the scale of the system remains unclear.
On software, the company says it has adapted Qwen, DeepSeek and Kimi models across different parameter sizes, precisions and long-context settings, and is working on a unified hardware and software stack to simplify deployment. Target uses include private enterprise knowledge bases with retrieval-augmented generation, office agents, financial compliance reasoning, offline government question answering and on-site industrial inspection, where local inference and data security matter.
The launch adds a system-level product to a growing set of Chinese RISC-V designs aimed at data centers, alongside recent server CPU and core IP announcements from other domestic developers. Whether the approach scales will depend on how chip performance, software ecosystem support and real deployments develop.
SEE ALSO: UltraRISC Unveils UR-CP400 RISC-V Core IP · SpacemiT V100 RISC-V Server CPU · DAMO Academy XuanTie E908A