The Pioneering Hardware Powering RF AI Workflows
Deepwave’s core mission is focused on providing easier access to RF Intelligence. This intelligence-first approach guides the development of Deepwave products, resulting in purpose-built RF software and hardware platforms that turn raw radio signal data into intelligence.
The first product Deepwave developed for generating RF intelligence was an integrated software and hardware edge platform. This was done out of necessity, as there were few solutions available to extract real-time intelligence from complex, data-heavy RF signals.
Deepwave launched its first offering for generating RF intelligence – the Artificial Intelligence Radio Transceiver (AIR-T). The patented AIR-T is the first software defined radio specifically for radio frequency and wireless deep learning.
How It Works
Deepwave’s hardware and software platform offers a new approach to analyzing RF data. Instead of relying on a network backhaul to stream raw RF (I/Q) data, Deepwave’s patented AIR-T hardware architecture and AirStack software allow AI models to analyze RF signals on the edge. This yields two sets of benefits.
The first is that low-latency ML model results can inform critical actions at the edge without requiring communication with an external network. In 5G networking, for example, RF intelligence at the edge can be used to inform spectrum sharing.
The second benefit of AI-generated intelligence at the edge is data reduction. Instead of attempting the technical feat of sending GBs per second of raw RF data across a network, the Deepwave platform only delivers ML classification results. This reduces data backhaul by a factor of more than 10 million. Converting RF signal data into RF intelligence at the antenna not only makes data transfers across networks more manageable, but it also can serve as timely intelligence to inform multi-sensor inputs to orchestration and command platforms.
Deepwave’s AIR-T hardware and AirStack software play vital roles in developing RF intelligence. A close up view of the AIR-T follows. On the software side, Edge, Core, and BitStream, all play their part to power high-performance computing on the AIR-T.
RF Intelligence Dataflow
The AIR-T’s orchestrated data path eliminates processing bottlenecks and is essential for developing a functional RF System of Action.
Capture and Pre-Process
(FPGA)
The RF transceiver captures wideband data. AirStack’s custom FPGA logic (customizable using AirStack BitStream) immediately performs high-rate pre-processing, data reduction, and feature extraction.
Transfer and Accelerate
(GPU/CPU)
The pre-processed data then moves via AirStack Core’s zero-copy mechanism to the shared GPU/CPU memory. The GPU executes complex AI model inference (e.g., signal classification) using accelerated libraries like TensorRT or CUDA.
Action and Management
(AirStack Edge/CPU)
Following AI model inference, the CPU runs the application logic, combining the AI-generated results with sensor metadata (like frequency, location, and reception time). AirStack Edge then securely publishes this low-bandwidth intelligence to external systems, or conversely, receives new commands to reconfigure the sensor for a new mission.
Modern Methods Require Modern SDR Hardware
The complexity and high data rates associated with modern radio frequency (RF) signals necessitate a departure from legacy signal processing methods. To address this requirement, Deepwave pioneered the integration of advanced computational power directly into the software defined radio (SDR) hardware. The integration of multiple processing architectures, known as heterogeneous computing, coupled with a patented Zero Copy memory management architecture, enable the AIR-T to perform high-performance computation exactly where the data is collected. This approach is essential for supporting demanding real-time workloads, such as continuous RF Intelligence and large-scale Electromagnetic Spectrum Monitoring.
Heterogeneous Computing
On the hardware side, Deepwave’s patented low-latency memory management architecture enables high-performance computing that leverages the AIR-T’s integrated heterogeneous computing hardware. This heterogeneous computing infrastructure consists multiple computing architectures:
Processor’s
Role
The central processing unit (CPU) manages the AirStack Core operating environment, handles general I/O, runs application logic for serial tasks, and coordinates the flow of final AI-generated RF intelligence.
The graphics processing unit (GPU) provides highly parallel processing capabilities, serving as the central engine for all Artificial Intelligence and Machine Learning operations on the platform. Deepwave’s products incorporate NVIDIA GPUs.
The field programmable gate array (FPGA) ensures strict real-time operations, ultra-low latency capability, hardware acceleration for signal processing, and direct data path management.
Processor’s Strength
The CPU provides the adaptable operating environment and executes the critical control plane, actively managing data movement among the RF frontend, the FPGA, and the GPU.
The GPU enables high-throughput operation for wideband signal processing and executes complex AI algorithms in real-time, delivering large-scale RF Intelligence.
The FPGA excels at front-end tasks like channelization, data reduction, signal conditioning, and pulsed signal analysis, offloading both the GPU and CPU.
Zero-Copy Architecture
Traditional SDRs connecting to discrete GPUs often suffer from prohibitive latency caused by multiple memory transfers (copying data from the radio to system memory, then to GPU memory).
Referring to the diagram below, the left side illustrates the conventional data pathway within a traditional SDR system. This architecture requires data to move from the RF front end through the FPGA and PCIe, necessitating an extra memory copy between distinct system memory and GPU memory pools. This extra copy introduces latency and bottlenecks the system.
In sharp contrast, the illustration on the right depicts the Deepwave AIR-T architecture, which eliminates this bottleneck. The Deepwave system utilizes a patented zero copy technology, where the GPU and CPU access a unified shared memory pool. This fundamental architectural difference enables direct data transfer from the RF front end to the processing units, resulting in a crucial latency reduction vital for real-time RF intelligence applications.
A New Approach to a Long-Standing Challenge
The AIR-T’s unique combination of heterogeneous computing and Zero Copy architecture is the driver for low-latency, low-bandwidth RF intelligence. Each computing architecture has its own strengths, but taken together, this combination provides the required high-performance compute.
Low Latency RF Intelligence
Achieve ultra-low latency RF intelligence and real-time system reaction by eliminating memory transfer bottlenecks through the utilization of a patented zero copy, unified memory architecture.
Resource Maximization and Efficiency
Maximize hardware utilization and computational efficiency by leveraging the field programmable gate array (FPGA) for critical, high-rate signal conditioning and data reduction. This directly frees the graphics processing unit (GPU) for complex artificial intelligence algorithms.
Scalable Network Advantage
Deliver mission-critical insights with superior network efficiency, converting high-volume raw RF data into actionable low-bandwidth intelligence directly at the edge. This process is engineered to cut network data backhaul by a factor of up to ten million.
Frequently Asked Questions
What are example use cases for Deepwave's SDR Hardware?
Example use cases include leveraging software platform integrations: the Radio Intelligence Agent (RIA) uses the AIR-T for performing real-time signal processing and speech recognition at the edge, securely delivering AI-generated data for complex automated workflows.
Additionally, Deepwave integrates with DeepSig’s OmniSIG Engine for high-throughput signal classification used in applications such as autonomous 5G network optimization and sophisticated Signal Analysis. This platform fundamentally increases operational tempo for deployment.
What core problem does the AIR-T solve in electromagnetic spectrum monitoring?
Deepwave’s AIR-T addresses the critical challenge of extracting real-time intelligence from high-volume, complex RF signal data. Traditional systems rely on network backhaul to stream gigabytes per second of raw data to central processing centers, which is slow and unmanageable – and ultimately results in historical data analysis. The AIR-T achieves real-time RF intelligence by moving AI model analysis directly to the edge, enabling immediate insights and actions.
How does the AIR-T's heterogeneous computing architecture prevent data bottlenecks?
The AIR-T integrates CPU, GPU, and FPGA hardware into a patented, low-latency memory management architecture. The FPGA is dedicated to high-rate signal conditioning and data reduction, offloading repetitive work. This frees the GPU to perform complex, parallel artificial intelligence (AI) inference in real-time, preventing the system from becoming bottlenecked.
What is the advantage of using a zero copy architecture on the AIR-T?
The zero copy architecture eliminates the crucial latency caused by multiple memory transfers required in traditional systems. Because the GPU and CPU access a unified shared memory pool, the RF data moves directly from the RF front end to the processing units, which is vital for real-time RF intelligence applications.
How does the AIR-T enable extreme network data reduction?
The AIR-T converts high-volume raw RF data into actionable, low-bandwidth ml classification results directly at the antenna. This innovative edge processing process substantially reduces data backhaul by a factor of up to ten million, transforming unmanageable raw data transfers into timely intelligence for orchestration platforms.
Which AirStack software component directly controls the FPGA for pre-processing?
AirStack BitStream is the dedicated AirStack software product responsible for customizing the FPGA. It provides the logic for the FPGA to immediately perform high-rate pre-processing, data reduction, and feature extraction, seamlessly starting the workflow for rf intelligence.