Sixteen Camera Lanes on a 45 by 30 Millimetre Module
Ezurio's Carbon AM67 puts TI's Sitara AM67x, 4 TOPS of deep learning accelerators, dual TSN Ethernet and four four-lane MIPI-CSI-2 inputs onto a solder-down OSM-MF module from about $338.
Ezurio has introduced the Carbon AM67, a 45 by 30 millimetre solder-down module in the OSM-MF v1.2 form factor built on Texas Instruments' Sitara AM67x. The processor carries up to four 64-bit Cortex-A53 cores at 1.4 GHz and three Cortex-R5F real-time cores at up to 800 MHz, two deep learning accelerators totalling up to 4 TOPS, and an IMG BXS-4-64 GPU rated to 50 GFLOPS with OpenGL ES 3.2 and Vulkan 1.2. Memory options are 4 or 8 GB of LPDDR4 with up to 128 GB of eMMC. Interfaces include two gigabit Ethernet ports with IEEE 1588 and TSN, PCIe Gen3 x1, USB 3.1 Gen1 and USB 2.0, four four-lane MIPI-CSI-2 camera inputs supporting up to sixteen virtual channels each, MIPI-DSI and LVDS driving three independent displays to 3840 by 1080 at 60 fps, plus CAN-FD, UART, I2C, SPI, PWM, GPIO, SDIO and I2S. Wi-Fi 6 or 6E with Bluetooth to 5.4 are options, and temperature grades are 0 to +70 °C commercial or −40 to +85 °C industrial. Pricing starts around $338 for 4 GB with 16 GB eMMC and reaches about $594 for 8 GB, with shipments scheduled for the fourth quarter of 2026.
The architecture is the story rather than the TOPS figure. Three Cortex-R5F real-time cores alongside the application processors means deterministic control and Linux can coexist on one device: the R5Fs run the motion loop, the safety logic or the protocol stack with hard timing, while the A53s run the application, the network stack and the inference. That removes a second microcontroller and, more importantly, removes the inter-processor communication link that is usually where the latency and the debugging difficulty live.
The camera capability is the other half. Four four-lane CSI-2 interfaces with sixteen virtual channels each is far more camera bandwidth than an inspection station needs and exactly what a multi-camera robot, an AMR or a 360-degree vehicle system needs. Combined with dual TSN-capable Ethernet, the intended application is clear: a vision-guided mobile machine that must also participate deterministically in a wider network.
Two notes for anyone evaluating it. The OSM-MF form factor is soldered down rather than socketed, which lowers cost and improves shock and vibration performance, and removes the option of swapping the module in the field — a real consideration in equipment that will be serviced rather than replaced. And 4 TOPS split across two accelerators is only useful if the toolchain can actually use both; TI's model compilation flow and operator support determine whether a given network reaches that figure or falls back to the CPU, which is a question to settle with the intended model during evaluation rather than after the board is laid out.