Modular RF Intelligent Processing Unit
Release date:
2026-05
Preface:
RFSoC+ Jetson AGX Orin is “ Integrate a high-end RF front-end, real-time signal processing, and AI-driven decision-making. “Integrated into a single edge‑system platform, its value lies not in isolated performance metrics, but in an end‑to‑end, low‑latency closed loop that spans from RF to intelligent decision‑making. In this issue, we’ll walk you through the system as a whole, then delve into NVIDIA Jetson and RFSOC, introducing this modular smart processing unit that combines RFSOC with a GPU.”
I. Overview
The GPU Orin+RFSoC heterogeneous intelligent processing module is designed for the edge. High-real-time, high-bandwidth, AI-intensive scenarios Its flagship hardware configuration leverages the NVIDIA Jetson Orin’s exceptional AI and general-purpose computing capabilities, combined with the AMD RFSoC’s high‑bandwidth direct RF sampling and hard‑real‑time signal processing capabilities, to build “ “A full‑stack heterogeneous platform for RF signal acquisition → real‑time processing → intelligent analysis → closed‑loop control” , suitable for high-end embedded applications such as software-defined radio, radar–communication integration, satellite communications, and intelligent sensing.
Figure 1: Physical image of the IW-RFSoC+GPU modular intelligent processing unit.
1.1 The Advantages and Capabilities of Combining the Two
RFSoC integrates an FPGA, ADCs/DACs, and RF functionality, enabling digital downconversion, filtering, FFT, beamforming, and other signal-processing tasks to be performed directly within the FPGA. The Jetson AGX Orin integrates a GPU, CPU, and Tensor Cores, enabling it to run deep learning, multi-sensor fusion, and decision‑making control. The RFSoC excels in signal processing but is weak in AI, whereas the Orin delivers strong AI performance but cannot directly process RF signals. By combining the two, it becomes possible to… Closing the loop from “signal → feature → AI → decision → control”
Figure: System Architecture Diagram of the IW-RFSOC+GPU Modular Intelligent Processing Unit
1.2 NVIDIA Jetson Orin
The Jetson AGX Orin was launched by NVIDIA in 2022. Embedded AI Computing Platform , positioned as an edge‑side “supercomputer,” with a primary focus on High performance, high integration, and configurable power consumption. , targeting high-end edge applications such as robotics, autonomous driving, industrial automation, and medical imaging
The AGX Orin is the flagship of the Jetson family, replacing the AGX Xavier. AI computing power has increased by approximately eightfold. , supporting the deployment of generative AI and multimodal large models. The Orin SoC is manufactured using an 8nm process and integrates a CPU, GPU, NVDLA, codecs, and high-speed I/O interfaces. Module dimensions: 100 × 87 mm , pin-compatible with the AGX Xavier, facilitating upgrades.
Figure 3: NVIDIA Jetson
|
NVIDIA Jetson Orin (AI Computing Core)
|
|||
|
AI computing power up to 275 TOPS (INT8) |
Up to 275 TOPS (INT8)
|
||
|
GPU architecture Ampere, 2,048 CUDA cores + 64 Tensor cores |
GPU Architecture: Ampere, 2048 CUDA cores + 64 Tensor cores
|
||
|
CPU |
12-core ARM Cortex-A78AE
|
||
|
Power consumption: configurable from 15 W to 60 W. |
|
||
1.3 AMD RFSoC
Figure 4: IW-AMD RFSoC (Radio Frequency and Real-Time Processing Core)
The IW-RFSoC-47DR algorithm verification and evaluation board provides a… COTS The solution leverages the advantages of RFSoC while more efficiently offloading data and capitalizing on the benefits of COTS: users achieve rapid deployment, lower total cost of ownership, robust lifecycle support, and simplified integration of future technologies. Zynq UltraScale+ devices integrate an ARM Cortex‑A53 processing subsystem, UltraScale+ programmable logic, and the highest signal‑processing bandwidth, enabling a comprehensive RF signal chain that meets the demanding requirements of high‑performance RF applications such as electronic warfare, quantum computing, and radar early warning.
|
IW-RFSoC-47DR (Radio Frequency and Real-Time Processing Core)
|
|||
|
PS side
|
4×DDR4 (4 GB, 64-bit, 2400 MT/s) |
||
|
2×QSPI flash firmware configuration file |
|||
|
PL side
|
4×DDR4 (4 GB, 64-bit, 366 MT/s) |
||
|
2×QSFP 100G optical ports |
|||
|
1×FMC+ expansion interface |
|||
·Data Interaction
|
High-speed interconnection |
Via PCIe Gen4 ×4/×8 |
·Analog-to-Digital Conversion (ADC)
|
Number of channels |
8 mmcx |
|
Sampling rate |
Up to 5 Gsps |
|
Quantization precision |
14bits |
|
Scope of Work |
0.05~6 GHz |
|
Input power |
Maximum input: 5 dBm |
|
Functional Mode |
Multi-channel synchronization, digital attenuation, real or complex sampling, quadrature correction, multi-band digital mixing and decimation, and other customizable i‑processing functions. |
·Digital-to-Analog Conversion (DAC)
|
Number of channels |
8 mmcx |
|
Sampling rate |
Up to 10 Gsps |
|
Quantization precision |
14Bbits |
|
Scope of Work |
0.05~6 GHz |
|
Output power |
Maximum output: -5 dBm |
|
Functional Mode |
Multi-channel synchronization, variable output power, real‑ or complex‑valued reconstruction, quadrature correction, multi‑band digital mixing and interpolation, and other customizable signal‑processing functions. |
IV. Core Product Advantages:
RF + AI + Integrated, End-to-End Low Latency
RFSoC direct RF sampling with microsecond‑level preprocessing, AGX Orin at millimeter‑level AI inference, End-to-end latency < 15 ms , meeting the real-time control requirements of radar, electronic warfare, wireless communications, and other applications.
Eliminates the multi‑stage conversion process of “RF → IF → Digital → AI,” Simplified links, reduced noise, and enhanced reliability
Ultra-high bandwidth plus massive memory, supporting concurrent operation of multiple sensors.
RFSoC supports 8/16 channels of 4 Gbps MIPI/PCIe 4.0, with a single lane delivering 100 Gbps, and can simultaneously interface with multiple phased-array radar channels+ IMU
AGX Orin: 64 GB LPDDR5, capable of caching 4 hours of 8K video (or 1 hour of radar point clouds), addressing the memory‑bandwidth bottleneck for large‑model weights and feature maps.
Extreme energy efficiency in heterogeneous computing, edge‑friendly deployment.
Total power consumption ranges from 30 to 80 W (RFSoC: 15–20 W + AGX Orin: 15–60 W), and is configurable. Its compute density far exceeds that of x86 servers—achieving the same level of performance would require over 300 W on an x86 platform.
Industrial-grade, wide-temperature operation and compact form factor, ideal for space-constrained, harsh-environment applications such as drones, vehicle-mounted systems, and airborne platforms.
A robust software ecosystem enables rapid development and deployment.
RFSoC supports Vitis, MATLAB/Simulink, and the RF Toolbox, enabling rapid verification of RF algorithms followed by FPGA implementation.
AGX Orin: Supports JetPack 6.X, NGC pre-trained models, the TAO fine-tuning toolkit, and ROS 2, enabling one-click deployment of AI models from training to the edge.
The two parties have jointly developed a co-simulation platform that supports FPGA–GPU collaboration, enabling parallel development of algorithms and hardware and reducing the development cycle by more than 50%.
Highly scalable, covering the entire lifecycle from prototyping to mass production.
Modular design: The AGX Orin is compatible with the Orin NX and Orin Nano, and its RFSoC features pin-to-pin compatibility, enabling seamless replacement and full‑range performance coverage.
V. Product Application Scenarios
Phased-array radar signal processing, electronic reconnaissance/jamming, Spectrum sensing 、
RFSOC: 8–16-channel radar signal acquisition, digital beamforming (DBF), moving target detection (MTD), and interference suppression.
AGX Orin: Radar‑based cloud‑AI target recognition (aircraft/UAV), Multi-radar data fusion Wait
Industrial Quality Inspection and Test & Measurement Field
The RFSOC multi-channel analog‑to‑digital transceiver, with features such as high‑speed data acquisition and synchronous triggering, can leverage its strengths in fields like image processing.
The AGX Orin can perform parallel processing of multi‑model inference and supports applications such as 3D modeling and industrial‑grade yield statistics.
Overall, the RFSOC handles signal transmission and reception and performs digital downconversion within the FPGA. It also implements filtering, FFT, beamforming, and low‑latency processing. The Jetson AGX Orin integrates a GPU, CPU, and Tensor Cores, enabling on‑board deep learning, multi‑sensor fusion, and decision‑making control.
VI. Comparison with Traditional Solutions
|
Comparison item |
RFSOC+AGX ORIN |
Traditional (FPGA + x86 server) |
Traditional (ASIC + Embedded CPU) |
|
AI computing power |
275 TOPS (INT8) |
100-200 TOPS (300W+) |
<10 TOPS |
|
End-to-end latency |
<15ms |
50-200ms |
>100ms |
|
Power consumption |
30-80W |
300-500W |
10–20W (low computing power) |
|
Volume |
100×87mm module |
19-inch chassis |
/ |
|
Multimodal fusion |
Native support |
Additional adaptation is required. |
Not supported |
|
Development cycle |
March–June |
June–December |
12–24 months |
Summary
RFSoC+ Jetson AGX Orin represents the optimal solution for today’s high-end edge‑based RF AI applications. Centered on “FPGA‑based hard‑real‑time RF processing combined with GPU‑powered, highly intelligent AI inference,” it delivers high bandwidth, low latency, powerful computational capabilities, low power consumption, a compact form factor, and ease of development—making it an ideal match for high‑value, highly complex, real‑time demanding use cases.
Keywords:
Contact Us
Phone:+86-29-84508304
Website: www.interwiser.com
Address: 25th Floor, Building E, Xidian Science & Technology Park, No. 168 Kechuang Road, Yanta District, Xi'an City
