IP cores for ultra-low power AI-enabled devices

Overview

Each nearbAI core is an ultra-low power neural processing unit (NPU) and comes with an optimizer / neural network compiler. It provides immediate visual and spatial feedback based on sensory inputs, which is a necessity for live augmentation of the human senses.

Optimized neural network inferencing for visual, spatial and other applications
Unparallelled flexibility: customized & optimized for the customer’s use case
Produces the most optimal NPU IP core for the customer’s use case: power, area, latency and memories trade-off
Minimized development & integration time

Ideal for battery-powered mobile, XR and IoT devices

Why nearbAI?

Highly computationally efficient and flexible NPUs

Enable lightweight devices with long battery life ... with ultra-low power, run heavily optimized AI-based functions locally
Enable truly immersive experiences ... achieve sensors-to-displays latency within the response time of the human senses
Enable smart and flexible capabilities ... fill the gap between “swiss-army knife” XR / AI mobile processor chips and limited-capability edge IoT / AI chips

Let's do a custom benchmark together:
provide us with your use case:

• Quantized or unquantized NN model(s):
ONNX, TensorFlow (Lite), PyTorch, or Keras

• Constraints:
Average power & energy per inference, silicon area, latency, memories, frame rate, image resolution, foundry + technology node

Key Features

L3 optimizer for optimal power, area and latency balance
Power: supports sub 1 mW always-on visual / spatial applications
NN compiler for firmware updates in the field
Configurable number of parallel MACs in convolution engine: 16 to 8192
Zero-latency NN switching - e.g., face ID NN gets faces only from face detection NN, multiplexed with other NNs
Integrates seamlessly with a wide range of RISC-V and ARM processor cores
Top computational efficiency for crystallized AI functions, yet programmable and field-upgradable
Typ. 4nm for mobile XR processor chip
Typ. 22nm for compact extreme edge AI chip
Configurable MAC accuracy: independent coefficient and data quantization, 4 to 16-bit, single-bit granularity
Supports model zoo of CNN / RNN / LSTM, and tailored customization
Long-term support: 5+ years

Benefits

Enable lightweight devices with long battery life
Enable truly immersive experiences
Enable smart and flexible capabilities

Block Diagram

Applications

Ideal for battery-powered mobile, XR and IoT devices

Deliverables

Standard off-the-shelf IP core(s)
Customized IP core(s)
IP core(s) + ASIC integration services
IP core(s) + full ASIC design services
Prototype

Technical Specifications

Foundry, Node

Independent

Availability

Available now

Request Info