Cambricon MLU 590

What it is

The MLU 590 is Cambricon's fourth-generation cloud AI chip, released in 2023 ^[3]. It is built on a 7 nm process and is the company's main chip for combined training and inference, aimed at models with hundreds of billions of parameters ^[1][2].

Training means running a model over large amounts of data and adjusting its internal weights until it behaves as intended; inference means using the finished model to answer real requests. A chip that handles both can serve both jobs inside one cluster, instead of forcing operators to buy one kind of machine for each stage. Parameters are the learned numbers inside a model, so a target of hundreds of billions of them says something about how much memory and memory bandwidth the hardware has to supply: the weights alone must fit somewhere and be moved constantly.

How it is built

Cambricon's hardware had reached its fifth-generation MLUarch microarchitecture with the MLU 590 ^[1]. A microarchitecture is the internal arrangement of a processor — how its compute units, memory paths and instruction set are organised — and a new generation usually means changes in how efficiently work is fed to those units rather than just more of them.

The chip supports chiplet heterogeneous integration and MLU-Link 8-card interconnect ^[1]. Chiplet integration means a processor is assembled from several smaller dies, each manufactured separately and packaged together, which allows different parts to be built on different process technologies. MLU-Link is the interconnect that lets eight cards exchange data inside one system, so that a group of chips can be treated as a single larger unit.

Compute is quoted at the cluster level: a cluster of MLU 590 chips delivers 2.048 PFLOPS in FP16 ^[1]. FP16 is a 16-bit floating-point format widely used in AI workloads because it halves the storage and bandwidth needed per number compared with 32-bit formats, at some cost in precision. When the chip was released, its compute parameters were benchmarked against Nvidia's A100 ^[3], which at the time was the reference point most buyers used to judge an accelerator.

Software

A chip is only usable if frameworks run on it. Cambricon's NeuWare software platform is designed so that code developed once can run across its chip lineup while remaining compatible with mainstream AI frameworks ^[1]. This matters as much as raw throughput: an accelerator that requires rewriting a training pipeline is expensive to adopt, and one that lets existing code move over with limited changes lowers the cost of trying it.

Release and deployment

According to the company's own timeline, the MLU 590 passed testing with internet customers and began shipping in 2024 ^[3]. ByteDance was reported to have deployed more than 100,000 MLU 590 and 690 chips in total ^[2]. A 2025 annual report review is the source for the FP16 cluster figure above ^[1]. Large-scale deployment by an internet company is a meaningful signal for this class of hardware, because such customers run sustained, heavy workloads rather than short trials.

Where it sits

Within Cambricon's lineup, the MLU 590 is the fourth-generation cloud part and the point at which the MLUarch design reached its fifth generation ^[1][3]. Its positioning — one chip for training and inference, aimed at very large models, with an interconnect for grouping cards and a software stack shared across the product family — describes the shape the company's cloud offering took: rather than selling isolated accelerators, it sells a chip meant to be clustered and programmed as part of a complete system.