Click to view the Featured Video

What is KPU™?

KPU™ is a domain-specific co-processor customized for specific applications, where “K” means “Kernel”, representing fundamental operations of many complex algorithms such as sort, CNN, rank, covariance, frequent pattern mining. KPU™ is designed for natively executing these kernels in ultra-high performance compared to its conventional CPU or GPU counterparts.

KPU™ is especially suitable for computing-intensive, streaming style, large data set, realtime applications. KPU™ is dedicated to enable a range of intelligent data analytics in Finance, Retail, Telecom, Energy, and Government, with the unprecedented computing capacity.


KPU™ is Software-Defined Accelerator

The slowdown of the Moore’s Law has rendered the conventional “tick-tock” model unsustainable. The disabled tandem of manufacturing technology (tick) and processor microarchitecture (tock) calls for more radical innovations in computer architectures. Domain-specific computing is regarded as a promising architecture evolution to uphold the expected performance growth rate in the post-Moore's Law era. The Software Defined Accelerator (SDA) is probably a viable approach to unlock the potential of the domain-specific computing paradigm.

SDA approach is an innovative methodology for designing and running Domain-specific Architectures. The essence of SDA is to bridge the semantic gap between domain-specific algorithms and micro-architectural IPs which are either commercially available such as various MCU cores and Memory blocks, or to be designed from scratch. SDA is also a runtime methodology which can help manage the accelerator resources more efficient. Specifically, all memory and cores are software-managed.


KPU™-equipped Systems

KPU™ penetrates the existing IT infrastructure to offload the computations that the current CPU or GPU cannot handle efficiently. Similar to GPU, KPU™ is built in PCIe devices to deliver computing power to the common CPU-centric servers.

We provide a series of products of KPU™ devices dedicated for specific niche market segments with specific computing patterns.

DataMonster : KPU™-powered OLAP Database

Database is a key infrastructure to support many data analytics applications. Many data analytics demand for massive query with realtime response that the state-of-the-art DBMS cannot provide.

We design a KPU™-powered in-memory DBMS, DataMonster, from ground up. DataMonster is customized for OLAP applications, compatible with SQL-92 standard. The user can use JDBC to connect DataMonster to executed high-throughput SQL data analytics.


Advanced Data Analytics


Time Series Analysis

Time series is prevalent in many data-driven applications. It’s a grand challenge to extract the information from the huge volume of time series data in realtime. KPU™ , designed with many built-in time series kernels, can solve the challenge in an elegant and, more importantly, in a cost-effective way.



AI & ML is a key approach to support effective decision-making processes in real world. However, most of the AI&ML algorithms is notorious for time-consuming, large data volume for model training and inference. KPU™ is customized for such computing & data intensive scenarios with advanced convolution, matrix, and regression kernels.


Complex SQL Query

SQL is a de facto standard for data operations. The conventional DBMS, however, strive to fulfill ACID but poor query executions. Moreover, some complex streaming algebras, such as percentile and sliding window, are not directly supported. KPU™ complements the conventional DBMS by supporting these complex query in hardware.


High Performance

KPU™ is the first kernel-based data processing engine which implement near-memory, data-driven, heterogeneous, and realtime features. Although KPU™ is not suitable for every application, it caters many computation-intensive ones such as computational finance, streaming database, and more.

Why KPU™?

KPU™ can fuel an innovative heterogeneous data processing paradigm. For the first time, the heterogamous applications get access to the customized heterogeneous computing infrastructures. GPU is designed for Graphic, while KPU™ family is designed for anywhere still bottlenecked in CPU and GPU world.

Easy to Use


High Performance

KPU™ Family