This company develops novel parallel neuromorphic hardware architecture combined with advanced algorithms to enhance AI and multivariate sensor processing. Their approach significantly improves Watt performance and reduces cost for edge computing applications. This is achieved by enabling numerous independent processing streams unconstrained by traditional monolithic memory structures.
Funding
Funding not disclosed
Founders
Product
Problem
Traditional computing architectures often struggle to efficiently process the complex algorithms required for modern AI and sensor applications, leading to bottlenecks and high power consumption. Existing GPUs, adapted from graphics processing, are not ideally suited for the evolving characteristics of AI software, particularly sparse and quantized models. This inefficiency limits the deployment of AI solutions, especially in edge computing environments with constrained power budgets.
Solution
Non-Von addresses these challenges with a novel parallel neuromorphic hardware architecture inspired by brain circuitry, coupled with cutting-edge algorithms. This approach eliminates the traditional memory bottleneck by pairing each core with its own memory, enabling a large number of independent processing streams. The resulting system significantly increases processing power while dramatically reducing electrical consumption, achieving an order-of-magnitude improvement in Watt performance. Non-Von's architecture is designed to natively support sparse and quantized AI models, optimizing performance for the latest efficiency-focused software trends.
Target Audience
The primary target audience includes AI developers, system integrators, and organizations deploying AI solutions in edge computing, autonomous systems, and data centers, particularly those requiring high performance with low power consumption.
Features
- Parallel neuromorphic architecture designed for AI from the ground up
- Chip architecture eliminates the traditional monolithic (von Neumann) memory bottleneck
- Native support for sparse, unstructured, and low-precision AI models
- Compatibility with common model development packages such as PyTorch, ONNX, Keras, and TensorFlow
- High throughput and low power consumption, suitable for edge computing applications
- SCULPT (Sparse Compact Ultra-Low Power Toolset) for AI automatically optimizes models for Non-Von hardware
- Demonstrated 10x performance improvement over Nvidia in MobileNet benchmarks