PerfXLab 澎峰科技

Discovering Potential. Driving Performance.

About Us

PerfXLab is an AI computing solution company. We provide performance products and services for embedded systems, cloud computing, as well as for artificial intelligence application needs.

Our team members come from MIT, Chinese Academy of Sciences, Intel, and NVIDIA, all of them being seasoned developers and researchers in performance optimization. We have led the development of OpenBLAS, as well as the OpenCL module in OpenCV.

We created InferXLite AI software stack for embedded systems. With PerfBox hardware, we empower our partners by embedded AI turn-key solution. For RISC-V community, we provide Perf-V creative board.


PerfBLAS is an optimized BLAS library for deep learning on ARM embedded system.

PerfBLAS can improve AlexNet model about 100% on ARMv7 and ARMv8 CPU cores.

PerfBLAS supports Linux and Android OS.


InferXLite is a lightweight deep learning inference framework for embedded systems. It supports ARM CPUs, ARM Mali GPUs, AMD APU SoCs, and NVIDIA GPU.

We provide the tool to covert DNN models (including Caffe model format and Darknet model format). In future, we will support other model formats.

InferXLite will be open source on github soon.


PerfBox is an ARM 64-bit embedded board. With InferXLite, and PerfBLAS AI software stack, we achieved squeezenet v1.1 model 87ms on single CPU core. PerfBox is used by smart camera, facial recognition system, new retail, and etc.


Perf-V is a FPGA demo board designed for RISC-V opensource community. It integrates various peripheral chips and offers many interfaces. Perf-V has great flexibility and transplant multiple architectures. We prepare various materials for you to learn out product and offer a perfect experiment platform for the design of RISC-V and FPGA products. It is a preferred hardware for you to study, develop programs and make a demo.

OpenBLAS Consulting Services

PerfXLab provides custom comsulting services and training for OpenBLAS. Our services include:

  • Deploying OpenBLAS
  • Port OpenBLAS on custom hardware or OS.
  • Developing wrappers or integrations with third party software
  • Performance tuning

Stay In Touch: