Introduction to the ESP32-S31 with Vision: Overview of the ESP32-S31 SoC

Understanding the capabilities of a new SoC

#

The ESP32-S31 is a high-performance system-on-chip for multimedia and AIoT applications with wireless connectivity. It combines two 32-bit RISC-V application cores, extensive wired and wireless connectivity, image and audio accelerators, and hardware security in one device.

This combination targets applications such as video doorbells, smart-home control panels, industrial automation, connected cameras, voice terminals, and appliances with rich human-machine interfaces. In this workshop, we focus on its ability to capture camera frames, run neural-network inference locally, and update a display without sending images to the cloud.

ESP32-S31 highlights

#

• Dual-core 32-bit RISC-V processor running at up to 320 MHz
• 128-bit SIMD data path for efficient AI and signal-processing workloads
• Dedicated hardware for camera input, display output, JPEG processing, pixel processing, and 2D data movement
• Wi-Fi 6, Bluetooth 5.4 LE and Classic, IEEE 802.15.4, and a Gigabit Ethernet MAC
• High-speed external flash and PSRAM interfaces with concurrent access
• Hardware security features including Secure Boot, memory encryption, protected key storage, and a Trusted Execution Environment

Processing and memory

#

Both application cores run at up to 320 MHz. One core has a 128-bit data path and SIMD instructions, allowing it to process multiple small integer values in parallel. Quantized neural networks rely heavily on multiply-accumulate operations over INT8 values, so this wide path is especially useful to ESP-DL.

The SoC also includes a 40 MHz low-power RISC-V core for work that does not require the application cores.

Camera frame buffers and model activations can be much larger than internal SRAM. The ESP32-S31-Korvo-1 therefore pairs the SoC with 16 MB flash and 16 MB PSRAM. Internal SRAM remains valuable for latency-sensitive data, while PSRAM holds large frames, display buffers, and model working memory.

Connectivity

#

The ESP32-S31 integrates several protocols that would otherwise require separate chips:

• 2.4 GHz Wi-Fi 6 (802.11ax): efficient wireless networking for connected devices
• Bluetooth 5.4 LE: LE Audio, Direction Finding, and Bluetooth Mesh 1.1
• Bluetooth Classic (BR/EDR): compatibility with established audio and HMI products
• IEEE 802.15.4: Thread and Zigbee, including Matter-over-Thread use cases
• 1000 Mbps Ethernet MAC: high-bandwidth wired networking with an external PHY

Vision inference does not require a network connection, but integrated networking makes it possible to send compact events—such as “safety zone entered”—instead of raw video.

Multimedia and new peripherals

#

The multimedia subsystem reduces the amount of pixel work performed by the CPU:

Other notable peripherals include:

• USB connectivity: USB 2.0 High-Speed OTG and USB Serial/JTAG
• Communication interfaces: CAN FD, SDIO, UART, SPI, I2C, and PARLIO
• Control: LED PWM
• Analog and sensing: Two 12-bit SAR ADCs, two DACs, a temperature sensor, and up to 14 capacitive-touch channels
• Audio: Audio Sample Rate Converter (ASRC) and dual I2S controllers with hardware-level Bluetooth audio synchronization
• Mathematical acceleration: CORDIC accelerator

Edge AI capabilities

#

An embedded vision pipeline performs more than inference. It must capture a frame, transform it into the model’s input format, run the model, decode its output, draw the result, and update the display. ESP32-S31 can distribute this work across dedicated blocks:

graph TD
    Camera[OV3660 or SC101IOT camera] -->|DVP frame| Memory[PSRAM frame buffers]
    Memory --> PPA[PPA preprocessing]
    PPA --> ESPDL[ESP-DL inference]
    ESPDL --> Results[Boxes and scores]
    Results --> Overlay[Display overlay]
    Memory --> Overlay
    Overlay --> LCD[4.3-inch LCD]

ESP-DL executes quantized .espdl models using target-optimized kernels. The SIMD core accelerates the neural-network operators, while the PPA and 2D-DMA can handle image preparation and display movement. This separation leaves more CPU time for application logic, networking, and user interaction.

Local inference offers:

• Low and predictable latency
• Operation without internet connectivity
• Reduced network bandwidth
• Better privacy because raw frames stay on the device

Security features

#

Vision devices often handle sensitive images and model intellectual property. ESP32-S31 provides multiple hardware security layers:

These blocks are building blocks, not an automatic security policy. A product must still configure Secure Boot, encryption, key provisioning, update validation, and access boundaries for its threat model.

What you learned

#

In this overview, you learned that the ESP32-S31:

• Combines dual 320 MHz RISC-V cores with a 128-bit SIMD data path
• Integrates Wi-Fi 6, Bluetooth 5.4 LE and Classic, IEEE 802.15.4, and a Gigabit Ethernet MAC
• Offloads camera, display, JPEG, and pixel-processing operations to dedicated hardware
• Runs quantized computer-vision models locally with ESP-DL
• Provides hardware mechanisms for verified boot, encrypted memory, protected keys, and software isolation

Next step

#

Overview of the ESP32-S31-Korvo-1

Return to the workshop main page