Ambient Scientific, a developer of programmable processors for implementing artificial intelligence (AI) capabilities in always-on edge and endpoint devices, has announced the beta release of software compilers for its GPX-10 processor.
It is making available two types of compilers – vertical compilers for fast development of AI applications, and a generic compiler to enable more programming flexibility. Developers will be able to download the vertical compliers along with the development environments for training, validation, and deployment to its GPX-10 AI processor; sample voice, vision and sensor-fusion applications are provided for quick proof of concept tests. Both the vertical and generic compilers are built to scale user applications across all ten of its programmable AI cores of the GPX-10 to optimize performance and power consumption.
The vertical compilers allow developers to create applications for specific applications like voice, computer vision and sensor fusion, using pre-determined neural networks. The vertical compiler for voice enables voice-based AI applications such as the creation of custom wake words, command phrase menu, and voice-IDs. This will then allow end users to create or change their wake-words, phrases, and voice-IDs as many times as they choose over the life of the product. This is particularly useful in situations where the device use-conditions are changed, or the user decides to repurpose an already deployed device.
A vertical compiler for vision enables the creation of various computer vision applications with up to 21 classes of vision-objects as defined by the OEM (original equipment manufacturer). The vertical compiler for sensor fusion enables developers to implement multi-parameter-based sensor fusion AI inferencing – which Ambient claims is a first of its kind in endpoint devices.
The company’s generic compiler enables developers to implement computer vision applications for recognition of up to 100 object classes. It supports most pre-trained and pre-built computer vision models (based on Mobilenet v1/v2, Resnet, VGG etc.) that are commonly used in industry for commercial or benchmarking purposes. Developers can also use Ambient Scientific’s generic compiler to develop and implement CNN based custom applications for computer vision such as object recognition and classification, presence and occupancy detection. These applications can be built with transfer learning from pretrained models or developed from scratch.
The generic compiler is framework agnostic. It allows developers to create neural network models using their own choice of AI frameworks such as TensorFlow, PyTorch, caffe, paddle-paddle, CNTK, C++, and deploy their models on GPX-10 AI processor.
The GPX-10 processor
Ambient Scientific was founded in 2017 with a purpose to develop software-defined AI processors for on-device inference and adaptive training to enable intelligence in always on and endpoint devices. It was founded by Gajendra Prasad Singh, the CEO, who spent his early career on SPARC development teams at Sun Microsystems, and prior to Ambient was VP of engineering at Wave Computing, where he worked on and invented various designs and implementations, including demonstrating a 10 GHz processor design using conventional CMOS processes.
At Ambient, he and his team have built their own processor architecture from the ground-up, to enable AI to be run effectively on an end device, which they call the DigAn. This architecture can scale from five cores to 8,000 cores and from 40nm technology to 7nm or 5nm technology, so the company is building a roadmap of AI processors to address a range of edge AI applications starting from endpoint AI devices, with the first products in the family of GPX AI processors being the GPX-10 and GPX-5.
For these processors, each core is software-defined and programmable, and provides support for multiple operand resolutions for both weights and inputs. Operand resolution can be set and modified dynamically by the application to optimize performance and power consumption. Each core can be enabled, disabled and halted dynamically; and character morphing between inference and training is possible, where the processors can switch modes between training and inference dynamically under application control.
The GPX-10 is an integrated SoC with 10 AI cores, 8 simultaneous sensor fusion, ultra-low-power ADCs, custom memories with ARM M4F core and peripherals. The company said it features a power consumption of 80 microwatts or less for always-on wake word or anomaly detection and can run always-on AI applications for multiple years on a single coin-cell battery. It adds that charger-less AI applications are possible by harvesting energy from various natural sources such as kinetic energy and RF energy.
- Software-defined vehicles drive next-gen connected EV development
- Selecting the right software defined radio solution for your application
- SOAFEE architecture for embedded edge enables software defined cars
- Software-defined Wi-Fi 6E AP triples Wi-Fi capacity and adds 6 GHz support