Neurxcore is offering the SNVDLA processor series, a Neural Processor Unit (NPU) product line based on an enhanced and extended version of the Open NVIDIA Deep Learning Accelerator. Depending on configuration, SNVDLA series are more than 100 times faster than a GPU for AI acceleration and 1000 times faster than a CPU. Silicon-proven, a typical implementation of our inference engine consumes sub-10mW of power consumption while running image classification tasks at movies frame rate. On the other hands, ultra-high performance (hundreds of TOPS) and concurrent processing are achievable thanks to his scalable and multi-core capable architecture. Fully configurable, it can be tuned to address specifics tasks, reaching the best accuracy/power/performance/area trade-off.


To efficiently run embedded AI applications on the SNVDLA, Neurxcore provides Heracium. Built upon the end-to-end and open-source Apache TVM framework, Heracium is a comprehensive set of software tools allowing you to explore AI system configurations, compile neural network models and check performances (accuracy/time execution). The neural network compiler comes with (and not limited to) the following features:

  • Supports multiple models: TensorFlow, Keras, Caffe, Caffe2, ONNX, PyTorch, mxnet, DL4J...
  • Runs everywhere: CPU, GPU, RTOS, Linux , bare metal
  • Offers heterogeneous execution across SNVDLA and CPU
  • Improves overall performance: optimized memory allocation, minimal runtime...


As a service, we can easily customize the SNVDLA and Heracium (introducing new operators or tuning the bandwith for instance). Neurxcore can provide you with optimized SNVDLA-powered subsystems or FPGA solutions for your specific application. These AI platforms can come with dedicated system memory, third-party core such as ARM Cortex or RISC-V, and high-speed bus such as PCIe. Neurxcore is offering extensive support on embedded software design, from firmware development to model training and optimization. Thanks to the versatility of our products, we can cover multiple AI use cases; from ultra-low power sensors and resource-limited embedded systems to more demanding and complex applications in the edge computing/datacenter field.