This document gives a hardware and software overview of the RED DSMC2 camera. The facts come from a SCARLET-W 5K on firmware 7.4.1.
- Application processor: a TI DM3725 (Cortex-A8, 1 GHz, 45nm) — confirmed from the
chip's top marking (
DM3725CBP100). It is in the OMAP3630 / DM37xx family (a DM3730 without the SGX graphics core, which this camera does not use — it has its own image coprocessors). 32-bit, one flat address space with no memory protection (no MMU). - Operating system: ThreadX (Express Logic). This real-time OS is now open source. The build uses the Green Hills compiler.
- Result of the flat memory: there is no process isolation. Any code that runs can read and write the whole memory.
- Memory map. Code and data run from DDR starting at
0x80000000. The application image isdsmc_program.mem(about 21.6 MB). To turn a virtual address into an offset inside that file, subtract the base:file_offset = VA − 0x80000000. The debug shell can read this memory live, and the range0x80000000–0x81FFFFFFis safe to read; reading unmapped or peripheral addresses faults the processor (see serial-shell.md).
Two architecture facts matter for anyone modifying the camera, and both are verified:
- The bootloader does not verify the installed OS image. It opens
/imageN/dsmc_program.mem, reads it, and jumps to it — no signature, no hash, no CRC, no decryption on that path. A patched image written into the active boot slot runs. - No memory protection. With a flat no-MMU address space, any code that runs is fully privileged and can read or write anything, including other cores' state.
Together these mean modified firmware does not need RED's signature to run on an owned body. RED's signing key is used only to gate the packaged upgrade flow, and the camera holds no signing private key. See keys.md.
The DSMC2 uses several processors. The main ARM controls a set of dedicated cores. Each core loads its own firmware image at boot.
| Subsystem | Core | Role |
|---|---|---|
| ISP (×2) | Tensilica Xtensa | Processes the sensor image (the R3D wavelet path is hardware) |
| IOB0 | Tensilica Xtensa | Handles SATA, HBA, and HDMI I/O (two SATA host controllers: HBA0 via the SATA3 FPGA, HBA1 direct) |
| IOB1 | Tensilica Xtensa | Does sensor dark-level and bad-pixel calibration |
| proio IOB0 | Tensilica Xtensa | Pro I/O build (data section is a zero-filled stub) |
| LSB FPGA | Xilinx Kintex-7 XC7K325T | Programmable logic; carries the SATA3 media path + a PI2DBS6212 2:1 SATA mux |
| SATA3 FPGA | on the LSB | Bridges the IOB's SATA up to SATA III; it is HBA0's register path (see the media folder) |
| DSP | TI TMS320C6x | Brain-module signal processing |
| WiFi/BT | RSI RS9113 (Redpine) | Wireless |
The ISP/IOB run Xtensa code loaded from NAND (config paths TENSILICA_*_BIN_PATH;
ISP_COUNT 2, IOB_COUNT 1). The ISA is confirmed by disassembly — the images load
cleanly as Xtensa in Ghidra. Whether the silicon is an ASIC or an FPGA is not
determinable from the firmware (it uses both words); only a physical method would
settle it. Load map, image layout, per-core function anchoring and the fabric
evidence are in peripheral-firmware.md.
Accessories carry their own MCU firmware. The rear modules, HDK, expanders and
JetPack are STM32 (confirmed in e.g. jetpack, hdk, portx, rvxl, rvexp,
vmexp); rcomPSOC is a Cypress PSoC; lens mounts, the LCD bridge (gw3400),
side handle, FlatUI and EVF each carry an image — see
peripheral-firmware.md and
accessory-bus.md.
The firmware uses standard, known parts:
- SQLite3 — the camera settings database (the CDB).
- zlib — on-NAND compression.
- YAFFS2 — the flash filesystem.
- LibTomCrypt — software crypto.
- TinyXML — the manifests.
- ProRes and DNxHD — the proxy and output codecs. The R3D wavelet is hardware; the ProRes proxy encode is a stage in the LSB FPGA (see whats-possible.md), not a CPU codec.
The internal NAND is a YAFFS2 flash filesystem split into partitions. The names
below are what nand ls shows on the device:
| Path | Role |
|---|---|
/env |
module/board identity: part and serial numbers, camera type |
/image1, /image2 |
the two OS image slots; one is active at a time |
/data |
user settings database and caches |
/log |
boot logs, run-time hours, boot count |
/pdata |
persistent data: calibration and the force_upgrade area |
/rodata |
large read-only data (per-model manifests, feature tables) |
User-facing folders also live on the filesystem: /presets, /looks, /luts,
/media, and /force_preset (where a preset is staged and consumed at boot).
There are three ways to reach the camera:
- WiFi — RCP control. The RED Command Protocol runs over TCP port 1111 (with a UDP discovery listener on 1112). RCP is documented in RED's own SDK and is the approved way to control and automate the camera. The same protocol also runs on the CTRL serial port (see the serial bridge).
- WiFi — file transfer (RFTP). RED's file transfer pulls user files
(
/presets,/looks,/luts,/log,/media, thumbnails) off the camera. This is the practical way to get bulk data off, including a firmware image that the shell has decompressed withnand zlib. - Serial — debug shell. A low-level interactive shell on the CTRL port, gated behind a preset flag. It gives live memory read/write, NAND access, the settings database, and hardware tests. See serial-shell.md. It is the only interface that reaches the diagnostics needed to repair a dead body.