SK-TDA4VM (TDA4VM)

Quick facts

Get it running

1. Backup: The SD card is the board's state and the image is its backup: tisdk-edgeai-image-j721e-evm.wic.xz (1.12 GB, SHA-256 7bebcff62930c93bc0d10462385a936305ffdfd2457c4bbd85c2b2f216076da0, from the SK-TDA4VM software page, no login). The box card was never read back, so there is no tested backup of its contents. Check a download with xz -t
2. Write the card: With a reader: tda4/tools/write-card.sh diskN path/to/image.wic.xz (refuses the internal disk, destroys the card). With none: the "No-reader path" in tda4/README.md
3. Boot: SW1 all OFF, open the console first, power on. Login root, no password, 115200 8N1
4. Check: The login: prompt. First boot 27.7 s, bootloader to login. python3 tda4/tools/warm_boot.py <n> times warm boots
5. First model: Models are in /opt/model_zoo, one folder each with artifacts and param.yaml. Push tda4/bench/zoo/tvm_bench.py with tda4/tools/push_serial.py and run it; the runtime log lists every node on the C7x (274 of 274 for the pose model)

You should see: a login: prompt on the serial console, 27.7 s after power-on on the first boot. Login is root with no password.

Every step, with the setup: the board's README.

What works

• CoreMark: 19,579.9 (GCC, Linux, -O3; 3 runs)
• Room-occupancy AI model, one inference: 0.0174 ms (GCC, -O3; 2000 MHz; 3 runs)
• Memory bandwidth, STREAM-style triad: 5,385 MB/s (GCC, Linux; 2000 MHz; 2 runs)
• Dhrystone 2.1: 5.263 DMIPS/MHz (GCC, Linux, -O3; 2000 MHz; 3 runs)

Download every result for this board (CSV), or compare it on this board on the Benchmarks page.

Projects and ports that ran here

• Run a small AI model on the chip: 20 of 20 test vectors right, plain C
• Measure how fast a chip runs: 9,835 CoreMark score at the board's top clock
• Run a vision model on a Linux board: 2.2 to 13.8 ms per model on the C7x accelerator
• A radar frame parser on the LP-AM13E230 and SK-TDA4VM Works: A radar frame parser in 882 bytes of C gives the same output on the AM13E230 and SK-TDA4VM as the Mac parser.
• An MSPM0 FFT example on the SK-TDA4VM Works: TI's MSPM0 FFT example runs unedited on the SK-TDA4VM and prints the same peak as on an R5F core and on a Mac.
• NVIDIA Holoscan's camera graph on the SK-TDA4VM Could work: NVIDIA Holoscan needs CUDA, so its camera graph is rebuilt from TI parts on the SK-TDA4VM: YOLOX-s pose takes 13.8 ms on the C7x accelerator.

Fixes

• If you see All four ports silent, a reset changes nothing, check SW1 and the port, then write the image.
• If you see Console silent after the USB cable moved, probe all four ports (tda4/tools/conlog.py) or set TDA4_PORT.
• If you see Shell stuck at a > prompt after a file push, wrap base64 at 76 columns; use tda4/tools/push_serial.py.
• If you see spl: ymodem err - Timed out during UART boot, use the current tda4/tools/uart_boot.py.
• If you see Failed to start Start psplash boot splash screen and Direct firmware load for j7-mcu-r5f0_1-fw failed with error -2, ignore both.
• If you see [ERROR] Error pulling frame from GST Pipeline, concatenate the clip to itself (20 copies).
• If you see ALG_PERR: ... e_afalg.c(463): Failed to bind socket : No such file or directory, use the A72 crypto instructions (AES-128-CBC 1.49 GB/s) or /dev/crypto.
• If you see 'FLAGS_STR' undeclared (CoreMark), add -DFLAGS_STR='"-O3"' and EEMBC's core_portme_posix_overrides.h.

Every fix for this board, with the exact error text

How long an AI assistant took here

For AI assistants

Point your assistant at this board's AGENTS.md (build, flash and fix notes) and the repository's llms.txt (the index).