French startup Neurxcore has launched a Neural Processor Units (NPU) AI chiplet built on an enhanced version of the open source Nvidia Deep Learning Accelerator technology.
The SNVDLA IP series from Neurxcore in Grenoble is aimed at image processing, including classification and object detection from wearables to the data centre. Depending on configuration, this can be more than 100 times faster than a GPU for AI acceleration and 1000 times faster than a CPU.
Neurxcore processors combine patented proprietary technologies with Nvidia’ open acceleration core. These neural processors are offered as licenses, which means that customers can use them in their own designs. Initially the Neurxcore processors will be available as chiplets that can be combined with CPU, memory and communications in a system in a package. This gives more flexibility by allowing the combination of different manufacturing processes, while optimizing costs in terms of density, integration and scalability says the company.
SNVDLA can also be used for generative AI applications and has already been proven in 22nm silicon from TSMC and shown on a demonstration board running a variety of applications. This can provide energy savings or approximately 10 times better performance compared to current equivalent commercial solutions through a patented in-memory calculation technology developed at the transistor level.
The IP package also includes the Heracium SDK (Software Development Kit) built by Neurxcore on the open-source Apache TVM (Tensor-Virtual Machine) framework. This is used to configure, optimize and compile neural network applications on SNVDLA products and check the accuracy and time execution. The compiler supports multiple models including TensorFlow, Keras, Caffe, Caffe2, ONNX, PyTorch, mxnet and DL4J and runs on the CPU, GPU, RTOS, Linux or bare metal as well as offering heterogeneous execution across SNVDLA and CPU
- Second-generation neuromorphic AI IP boosts efficiency
- The AI chips race is on – what role will IP play?
- Quadric joins Silicon Catalyst to offer AI IP
The product line caters to a wide range of industries and applications, spanning from ultra-low power to high-performance scenarios, including sensors and IoT, wearables, smartphones, smart homes, surveillance, Set-Top Box and Digital TV (STB/DTV), smart TV, robotics, edge computing, AR/VR, ADAS, servers and more.
Neurxcore offers a complete package allowing the development of customized NPU solutions, including new operators, AI-enabled optimized subsystem design, and optimized model development, covering training and quantization.
“80% of AI computational tasks involve inference. Achieving energy and cost reduction while maintaining performance is crucial,” said Virgile Javerliac, founder and CEO of Neurxcore.
The IP has fine-grain tunable capabilities, such as the number of cores and multiply-accumulate (MAC) operations per core, to allow for versatile applications across diverse markets. The inference stage, which involves using AI models to make predictions or generate content, is a pivotal aspect of AI. Neurxcore’s design addresses this phase efficiently, making it suitable for various applications, even when serving multiple users simultaneously.