The SoC is fabricated on a 5 nm EUV process, delivering a for typical vision pipelines (e.g., 30 fps @ 1080p object detection). The modular design enables scaling the NPU and ISP blocks independently, allowing OEMs to tailor the chip to cost‑sensitive or performance‑critical applications.
| | Description | |-----------|-----------------| | CPU | 4‑core ARM Cortex‑A78AE, 2 GHz, with hardware virtualization for secure multi‑tenant workloads. | | NPU | 2‑stage neural‑processing unit (NPU) – a vector‑core (V‑core) for high‑throughput FP16/INT8 ops and a tensor‑core (T‑core) optimized for depth‑wise convolutions and transformer attention heads. | | ISP | 12‑bit, 4‑lane MIPI CSI‑2 ISP supporting up to 4 MP (3840 × 2160) @ 60 fps RAW capture, with on‑chip HDR, noise‑reduction, and 3A (auto‑exposure, auto‑focus, auto‑white‑balance) pipelines. | | DSP | Fixed‑function audio/video codecs (H.264, H.265, AV1) and a low‑latency audio DSP for beam‑forming microphones. | | Memory | Up to 8 GB LPDDR5X (6400 MT/s) + 256 MB on‑chip SRAM. | | Security | Secure boot, hardware root of trust, on‑chip crypto engine (AES‑256, SHA‑3). | | Interfaces | 2× MIPI‑CSI, 2× MIPI‑DSI, 1× HDMI 2.1, 2× USB‑3.2, 2× PCIe Gen 3 (x2), 1× Gigabit Ethernet, CAN, I²C, SPI, GPIO. |
: Unlike static scans, these are video sequences captured using mobile phones. This introduces "in-the-wild" artifacts such as: Variable lighting and glare. Motion blur and perspective distortion. Complex backgrounds and hand occlusions.
The SoC is fabricated on a 5 nm EUV process, delivering a for typical vision pipelines (e.g., 30 fps @ 1080p object detection). The modular design enables scaling the NPU and ISP blocks independently, allowing OEMs to tailor the chip to cost‑sensitive or performance‑critical applications.
| | Description | |-----------|-----------------| | CPU | 4‑core ARM Cortex‑A78AE, 2 GHz, with hardware virtualization for secure multi‑tenant workloads. | | NPU | 2‑stage neural‑processing unit (NPU) – a vector‑core (V‑core) for high‑throughput FP16/INT8 ops and a tensor‑core (T‑core) optimized for depth‑wise convolutions and transformer attention heads. | | ISP | 12‑bit, 4‑lane MIPI CSI‑2 ISP supporting up to 4 MP (3840 × 2160) @ 60 fps RAW capture, with on‑chip HDR, noise‑reduction, and 3A (auto‑exposure, auto‑focus, auto‑white‑balance) pipelines. | | DSP | Fixed‑function audio/video codecs (H.264, H.265, AV1) and a low‑latency audio DSP for beam‑forming microphones. | | Memory | Up to 8 GB LPDDR5X (6400 MT/s) + 256 MB on‑chip SRAM. | | Security | Secure boot, hardware root of trust, on‑chip crypto engine (AES‑256, SHA‑3). | | Interfaces | 2× MIPI‑CSI, 2× MIPI‑DSI, 1× HDMI 2.1, 2× USB‑3.2, 2× PCIe Gen 3 (x2), 1× Gigabit Ethernet, CAN, I²C, SPI, GPIO. | midv276
: Unlike static scans, these are video sequences captured using mobile phones. This introduces "in-the-wild" artifacts such as: Variable lighting and glare. Motion blur and perspective distortion. Complex backgrounds and hand occlusions. The SoC is fabricated on a 5 nm
