The Orange Pi 4 Pro is an Orange Pi single-board computer built around the Allwinner A733 SoC. It pairs two Cortex-A76 cores at 2.0 GHz with six Cortex-A55 cores at 1.8 GHz, an Imagination PowerVR BXM-4-64 GPU, a VeriSilicon NPU rated at up to 3 TOPS, hardware video encode and decode, and LPDDR5 memory in configurations from 4 to 16 GB. Launch pricing started at $35 (€32) for the 4 GB model. The 6 GB unit tested here launched at $40 (€37).
The A733 puts this board in an interesting position. Its multi-core CPU throughput nearly matches the Raspberry Pi 5, the GPU and NPU both work out of the box on the vendor image, and the hardware video engine handles 4K HEVC decode and transcode in real time. The software stack is still maturing, and the hardware underneath is capable for the price.
Specifications
- SoC: Allwinner A733 (2x Cortex-A76 @ 2.0 GHz, 6x Cortex-A55 @ 1.8 GHz, XuanTie E902 RISC-V co-processor)
- GPU: Imagination PowerVR BXM-4-64 MC1 (GLES 3.2, Vulkan 1.3, OpenCL 3.0)
- NPU: VeriSilicon VIP, up to 3 TOPS (INT8/INT16/FP16/BF16)
- RAM: 4/6/8/12/16 GB LPDDR5 (6 GB as tested)
- Storage: microSD, eMMC module socket (also UFS-capable), M.2 M-key PCIe 3.0 x1, 16 MB SPI NOR flash
- Networking: Gigabit Ethernet (with PoE support), Wi-Fi 6 (AIC8800D80, 2.4 + 5 GHz, 1x1), Bluetooth 5.4
- Display: HDMI 2.0 (up to 4K at 60 Hz), MIPI DSI
- Camera: Dual MIPI CSI (1x 4-lane, 1x 2-lane)
- USB: 1x USB 3.0 Type-A, 3x USB 2.0 Type-A, 1x USB-C (power)
- GPIO: 40-pin Raspberry Pi-compatible header (I2C, SPI, UART, PWM)
- Other: Onboard RTC (HYM8563), IR receiver, 3.5 mm audio jack
- Board size: 89 x 56 mm (3.5 x 2.2 in), 58 g (2.0 oz)
- Power: 5 V / 3 A USB-C
- Launch price: $35/€32 (4 GB), $40/€37 (6 GB), $45/€41 (8 GB), $55/€50 (12 GB)
Design and Build
The board follows the standard SBC credit-card form factor at 89 x 56 mm (3.5 x 2.2 in). Full-size HDMI sits on one edge, four USB Type-A ports (one blue USB 3.0, three USB 2.0) and Gigabit Ethernet along the other, with the USB-C power input and 3.5 mm audio jack on a third side. The 40-pin GPIO header sits along the board's long edge in the familiar Raspberry Pi position.
On the underside, there is a removable eMMC module socket that also accepts UFS modules, and an M.2 M-key slot for NVMe SSDs on a PCIe 3.0 x1 lane with a link capability of 8 GT/s. The unit tested had a 58 GB eMMC module fitted but no NVMe drive. A 16 MB SPI NOR flash chip is present for bootloader storage.
The board includes an onboard HYM8563 RTC that works once set from the system clock, though it comes up uninitialized on a fresh boot without battery backup. Dual MIPI CSI connectors support up to two camera modules, and PoE headers are present for powered Ethernet deployments. The 40-pin header is mapped via Orange Pi's wiringOP tool and exposes I2C, UART, SPI, and PWM. SPI and additional serial ports are enabled through device tree overlays and a reboot; the vendor ships overlays for all of the header's alternate functions.
Performance
CPU and Memory
The A733's two Cortex-A76 cores at 2.0 GHz and six Cortex-A55 cores at 1.8 GHz deliver a 7-Zip multi-core score of 10269 MIPS, with a single-core result of 2305 MIPS on the A76 cluster. For context, the Raspberry Pi 5 with four Cortex-A76 cores at 2.4 GHz scores around 10930 MIPS multi-core and 3136 MIPS single-core in sbc-bench. The Orange Pi 4 Pro reaches about 94% of the Pi 5's multi-core throughput by complementing its two big cores with six small ones. The single-core result is about 73% of the Pi 5's, reflecting the lower A76 clock speed.
Memory bandwidth measured at 4222.8 MB/s (memcpy) and 7750.6 MB/s (memset) on the A76 cores, and 2763.2 / 7691.5 MB/s on the A55 cluster. These figures can vary between runs on the same boot; an earlier measurement on the same session recorded 5203.9 MB/s memcpy on the A76 cores, likely due to DRAM frequency-scaling state that cannot be read on this image. The board carries 6 GB of LPDDR5 in this configuration.
AES-256-CBC throughput on the A76 cores reached 1141408 KB/s, driven by the ARMv8 cryptographic extensions (AES, SHA1, SHA2, CRC32). The A55 cluster managed 838074 KB/s on the same test.
Storage
The board's eMMC module (58 GB, MMC 5.1) delivered 39.2 MiB/s random read (10024 IOPS) and 38.5 MiB/s random write (9858 IOPS) at 4K with a queue depth of 32, and sequential throughput of about 45 MiB/s in both directions. These figures reflect the module running in high-speed 52 MHz mode (8-bit, 3.3 V), not the HS200 or HS400 modes the module supports.
The M.2 PCIe 3.0 x1 slot reports a link capability of 8 GT/s x1. No NVMe drive was fitted during testing.
GPU
The PowerVR BXM-4-64 MC1 GPU is driven by Imagination's DDK 24.2 stack, preinstalled on the vendor image. It supports OpenGL ES 3.2, Vulkan 1.3 (apiVersion 1.3.277, above the 1.2 on the spec sheet), and OpenCL 3.0 out of the box.
glmark2-es2 at 1920x1080 (off-screen, via a bare X server) scored 273. vkmark in headless mode at the same resolution returned 429. Desktop OpenGL (GLX) falls back to software rendering (llvmpipe), since the vendor stack provides GLES/EGL and Vulkan rather than a desktop GL path.
On the compute side, clpeak measured 62.18 GFLOPS at FP32 and 62.34 GFLOPS at FP16 (no 2x FP16 rate), with global memory bandwidth of 5.13 GB/s. The GPU ran at its top OPP of 1008 MHz during these tests. OpenCL is functional and useful for light image processing, though the Vulkan driver's 16 KiB shared memory limit currently prevents llama.cpp from offloading inference to the GPU.
Video Decode and Encode
The A733's cedarc video processing unit handles encode and decode through GStreamer's OpenMAX elements. This BSP does not expose V4L2 m2m nodes, so ffmpeg, mpv, and most browsers do not use the hardware path on this image.
Decode throughput across codecs and resolutions (300 frames each, decoded to fakesink with no display sync):
- HEVC 1080p: 256.0 fps
- HEVC 2160p: 64.8 fps
- HEVC Main10 2160p: 56.4 fps (150 frames)
- VP9 1080p: 210.1 fps
- VP9 2160p: 60.9 fps
- H.264 1080p: 130.5 fps
- H.264 2160p: 30.7 fps
- MJPEG 1080p: 100.5 fps
HEVC and VP9 at 4K decode with roughly 2x real-time headroom, and 10-bit HEVC Main10 at 4K also works. H.264 at 4K reaches 30.7 fps, enough for 4K30 content but not 4K60. 8K HEVC and VP9 decode, a headline SoC specification, is not yet accepted by the cedarc library on this image.
Hardware transcoding (decode then re-encode with zero-copy DMA between decoder and encoder) ran at 86.6 fps for HEVC 1080p to H.264, and 30.6 fps for HEVC 2160p to H.264. The encoder and decoder are separate hardware blocks (VE and VE2), so both operate simultaneously. 4K30 HEVC-to-H.264 transcode runs in real time.
Playing back 8-bit HEVC 2160p30 to the HDMI output via kmssink, the board delivered 30.07 fps with zero dropped frames and 30% CPU usage. H.264 at the same resolution dropped 48 out of 300 frames, consistent with the decode rate sitting right at the 30 fps threshold. Playing 10-bit HEVC 2160p to the display triggered an IOMMU fault in the video decoder on this BSP, and hardware decoding was unavailable until a reboot.
Multi-stream decode saturated at about 140 fps aggregate for 1080p H.264, enough for roughly four simultaneous 1080p30 camera streams. Two 4K30 HEVC streams ran at about 55 fps aggregate, so the hardware can service two live 4K HEVC feeds.
AI Inference
Running Llama 3.2 1B (Q4_K_M) on the CPU with all eight threads, the board managed 42.7 tokens per second for prompt processing (pp512) and 6.8 tokens per second for text generation (tg128), averaged over five runs with under 0.6% standard deviation. GPU offload via Vulkan is not possible on this image due to the shared memory limit.
Core placement matters on this asymmetric CPU. Token generation on just the two A76 cores reached 7.97 tok/s, faster than using all eight cores (6.80 tok/s), because mixing in the slower A55 cores reduces generation speed. Prompt processing, which parallelizes more effectively, peaks at 46.58 tok/s with all eight cores.
The VeriSilicon NPU runs at 852 MHz in the image's default performance governor. The vendor ships a YOLOv5 demo and benchmark tool on the image, and the demo correctly detected objects in the sample image (dog 86%, truck 65%, bicycle 56%). The benchmark ran 640x640 object detection at an average of 47.98 ms per frame (20.8 FPS) over 100 iterations. About 18 ms of each frame is CPU pre- and post-processing running on an A76 core; on a single A55 core, that CPU portion rises to about 79 ms, giving 8.9 FPS. The NPU inference itself takes about 3 ms.
Running the NPU vision pipeline and llama.cpp at the same time without core pinning reduced both: LLM generation dropped from 6.8 to 3.46 tok/s, and YOLO fell to 13.2 FPS. Pinning the LLM to the two A76 cores and YOLO to the six A55 cores restored LLM generation to 6.88 tok/s, at the cost of YOLO throughput (8.0 FPS, limited by A55 pre/post-processing speed). On this board, the two big cores are the key resource in any AI pipeline, and choosing how to allocate them is the central scheduling decision.
Connectivity and I/O
The single USB 3.0 Type-A port sits on an xHCI controller with USB 3.1 Gen 1 (5 Gbps) root hub capability. The three USB 2.0 Type-A ports are served by an onboard Terminus four-port hub on a separate EHCI controller.
Gigabit Ethernet is present via the dwmac-sunxi MAC. Network throughput was not measured in this test run.
Wi-Fi uses the AIC8800D80 chip on SDIO, supporting 802.11ax (Wi-Fi 6) on both 2.4 GHz and 5 GHz bands with HE80 channels on 5 GHz. It is a single spatial stream configuration. Bluetooth 5.4 is available over UART. The Bluetooth hardware address on this unit was a default value (10:11:12:13:14:15) rather than a unique address, so users running multiple boards on the same site should assign unique addresses.
HDMI 2.0 output drove a 4K display at 60 Hz. The kernel log showed some HDMI audio errors and periodic hot-plug/replug events from the attached monitor.
Software and Linux Support
The tested image is "Orange Pi 1.1.0 Resolute," based on Ubuntu 26.04 with an XFCE desktop. It runs Orange Pi's vendor BSP kernel 6.6.98-sun60iw2, built on 2026-07-31, with a vendor U-Boot 2018.07 bootloader.
The kernel uses four out-of-tree modules: pvrsrvkm for the PowerVR GPU, vipcore for the NPU, sunxi_ve for the video engine, and aic8800 for Wi-Fi. The board's specialized hardware is supported through these vendor drivers.
Mainline Linux support for the A733 is in progress. Clock controller support (CCU/R-CCU) has landed upstream, and the linux-sunxi community is working on further integration. A community Armbian build using mainline 7.1.5 exists, though the board is not yet an official Armbian target. Mainline builds do not yet include GPU acceleration, NPU, or hardware video support.
The vendor image ships the GPU driver stack and NPU runtime preinstalled. The YOLOv5 demo tools in /opt/yolov5 come bundled with the image. Hardware video is accessed through GStreamer's cedarc OpenMAX elements; without V4L2 m2m nodes, ffmpeg does not use the hardware encoder or decoder. This determines which media applications can take advantage of the VPU today.
The Wi-Fi driver prints scan debug messages to the kernel log every couple of minutes, and the RTC logs warnings until it is manually initialized.
Thermals, Power and Noise
Under a 10-minute all-core stress-ng load, both CPU clusters held their maximum clocks (1794 MHz on the A55, 2002 MHz on the A76) for the entire run with no throttling. Idle temperatures started at 28.6 to 29.6°C (83.5 to 85.3°F) across all five thermal zones. Under load, the A55 cluster zone peaked at 65.2°C (149.4°F) and the A76 zone at 64.3°C (147.7°F), with DDR at 57.2°C (135.0°F) and the NPU zone at 55.7°C (132.3°F). The CPU's first passive throttling trip is set at 90°C (194°F).
This was not the hottest workload. Running llama.cpp on all eight threads pushed the A55 cluster zone to 77.2°C (171.0°F) within about 30 seconds, still below the 90°C (194°F) trip but notably warmer than a stress-ng integer load. Cooling counters showed no throttling during any measured test phase. Sustained NPU inference over 3000 iterations peaked at just 48.8°C (119.8°F) on the NPU zone, making vision workloads thermally easy by comparison.
The cooling setup on the test unit was not recorded. The board has no fan or active cooling managed in software.
Everyday Use
The board boots from microSD in about 18 seconds to an XFCE desktop.
As a 4K media player, the board handles 8-bit HEVC content at 4K30 with zero dropped frames when driven through GStreamer to the HDMI output. The practical requirement is using GStreamer-based pipelines or a player built on them, since ffmpeg-based tools do not use the hardware decoder on this BSP. HEVC and VP9 sources are the better choice for 4K; H.264 4K decode runs at the 30 fps threshold with no headroom.
For edge AI with a camera, the NPU runs YOLOv5 at about 20 FPS when an A76 core handles the CPU-side processing. In a split configuration with the LLM on the two A76 cores and vision on the six A55 cores, the board manages about 7 tok/s text generation alongside 8 FPS object detection simultaneously. The two big cores are the main constraint in any combined AI pipeline.
The eight cores and 6 GB of LPDDR5 make the board comfortable for running multiple containers as a home server, DNS box, or self-hosted application host. For a small NVR, the video decoder handles about four simultaneous 1080p30 H.264 streams or two 4K30 HEVC streams. Adding an NVMe SSD via the M.2 slot would help with storage-heavy workloads, bypassing the eMMC speed limitation.
Verdict
The Orange Pi 4 Pro delivers eight-core CPU performance close to the Raspberry Pi 5, a working GPU with Vulkan and GLES, functional hardware video encode and decode, and an NPU with preinstalled tools, at a launch price of $35 (€32). Its media capabilities are strong for the price: real-time 4K HEVC decode and transcode, multi-stream camera decode, and a functioning AI inference pipeline with the vendor's YOLOv5 demo ready to run.
The software stack is still maturing. The vendor BSP kernel is the practical option today, with mainline support in progress. The eMMC currently runs in 52 MHz high-speed mode rather than its HS200/HS400 modes. The video engine is accessed through GStreamer's OpenMAX path rather than V4L2, so ffmpeg and most Linux media players do not use it. And 10-bit HEVC playback to the display triggered a decoder fault on the test image. These are software-layer issues that updates can address; the underlying hardware is not at fault.
For a self-hosting box, a GStreamer-based media player or transcoder, or an edge AI platform where the NPU and camera interfaces matter, the Orange Pi 4 Pro is a strong option at its price. Those who need a polished desktop experience or depend on ffmpeg for media workflows should verify whether the software support has progressed beyond the current image before committing.
