Kunlunxin P800

What the P800 is

The Kunlunxin P800 is an AI accelerator — a processor designed for the matrix and vector arithmetic that neural networks are built out of — and it is the third generation of Kunlunxin's accelerator line. It was launched in 2024^[1][2]. The chip is built on XPU-P, an architecture Kunlunxin developed in-house rather than licensing from another vendor^[1].

Accelerators of this kind sit in servers alongside ordinary CPUs: the CPU handles scheduling, storage and data movement, while the accelerator does the bulk of the arithmetic. How fast that arithmetic runs, and how much data it can keep close at hand, is what determines how large a model can be trained or served on a given machine.

Specifications

Kunlunxin rates the P800 at 345 TFLOPS in FP16^[1][2]. TFLOPS here means trillions of floating-point operations per second, and FP16 means the numbers being multiplied are held in 16-bit floating point. The card carries 96 GB of high-bandwidth memory (HBM) and supports FP8, FP16 and FP32 precisions^[2].

Precision is a trade-off that shows up throughout AI computing. Fewer bits per number means each value takes less memory and the arithmetic units can move more of them per second, so throughput rises; more bits preserve detail and dynamic range, which training is usually more sensitive to than inference. A chip that supports FP8 through FP32 can be used across that range instead of being tied to one format.

Memory capacity matters as much as raw speed. Model weights, activations and the key-value cache that grows with context length all have to fit in or stream through the card's memory; 96 GB of HBM is the budget a single P800 has to work with^[2].

Systems built on it

The P800 is used both as a single card and as a building block for larger machines. In April 2025 Baidu lit up a 32,000-card supercluster based on the P800 that can carry multiple training jobs for models with hundreds of billions of parameters at once^[1][2]. A cluster of that size matters because training at that scale requires the work to be split across many chips and kept in constant communication; the interconnect and scheduling around the accelerator are as important as the chip itself.

At the smaller end, Kunlunxin said a single server with eight P800 cards reached a throughput of 2,437 tokens per second^[1]. Tokens are the units of text a language model reads and produces, so tokens per second is a direct measure of how much generation a server can sustain.

In April 2026 a 256-card Tianchi supernode built on the P800 was powered on, with 25% higher throughput and 50% better inference efficiency than the previous generation^[2]. A supernode packages many accelerators into one tightly connected unit, which reduces the cost of moving data between them.

Software support and customers

Hardware is only usable if models run on it. The P800 has been adapted to ERNIE, DeepSeek, GLM and MiniMax models and deployed at China Merchants Bank, China Mobile and State Grid^[2]. Those deployments put the card in banking, telecommunications and power-grid settings, where workloads range from serving chat models to the pattern-recognition and forecasting tasks those industries run.

In August 2025 Kunlunxin won a China Mobile AI server procurement worth more than 1 billion yuan, taking 70% and 100% shares in several CUDA-compatible segments^[1]. CUDA is Nvidia's programming model, and code written for it is the main reason customers stay with one vendor; a segment described as CUDA-compatible is one where the bidder offers a path for existing CUDA-based software to run with limited changes.

Where it sits

The P800 is a domestic accelerator aimed at both training and inference at scale, and the evidence of its position is less in a spec sheet than in the systems around it: a 32,000-card cluster, an eight-card server throughput figure, a 256-card supernode, framework support for several widely used model families, and procurement wins measured in fractions of a bidding segment. Those are the pieces that decide whether a chip is a demonstration or a product. The P800's record as of the facts above is that it has moved into volume clusters and customer sites, with each generation of systems reporting gains over the one before it^[1][2].