TH1520 Physical Board Procurement And Test Plan

This plan selects three TH1520 development boards as physical software baselines for Eliza-AI-SoC: BeagleV-Ahead, Sipeed Lichee Pi 4A, and Milk-V Meles. The goal is not to claim phone-class performance. The goal is to get reproducible Linux and Android-adjacent boot, debug, log, and benchmark evidence on real quad-core C910 silicon with a TH1520 NPU/GPU/display stack.

Board Set

Buy two of each board: one clean reference unit and one destructive/recovery unit. Do not mix board revisions in a single result table unless the revision is part of the result key.

Required Accessories

Minimum shared lab kit:

• Linux host with fastboot, adb, tio, picocom, minicom, screen, git, gcc, make, fio, lmbench, stress-ng, perf, bc, jq, usbutils, and pciutils where available.
• USB power meter capable of logging 5 V input power.
• Powered USB hub with per-port power switches.
• Gigabit Ethernet switch and known-good CAT6 cables.
• HDMI monitor, HDMI capture dongle, and the board-specific HDMI cable: mini-HDMI for BeagleV-Ahead, full-size HDMI for Meles, and the connector required by the purchased Lichee Pi 4A kit.
• Class A2 or better 64 GB microSD cards, at least two per board.
• USB-to-UART adapters with selectable or verified logic level. Use CP2102N or Raspberry Pi Debug Probe for BeagleV-Ahead. Use Sipeed RV Debugger Plus or a verified 3.3 V-compatible adapter for current Lichee Pi 4A release boards.
• JTAG/debug tools: Sipeed RV Debugger Plus or SLogic Combo 8 in CKLink mode for Lichee Pi 4A, and a RISC-V-capable JTAG adapter plus board-specific cable for BeagleV-Ahead and Meles.
• ESD mat, ESD wrist strap, Dupont leads, spare USB-C data cables, USB 3.0 micro-B data cable for BeagleV-Ahead flashing, and USB-A-to-C data cable for Meles eMMC flashing.

Board-specific accessories:

Image And Flashing Paths

Record every image as artifacts/th1520/<board>/images/<image-name>/ with:

sha256sum * | tee SHA256SUMS
date -u +"%Y-%m-%dT%H:%M:%SZ" | tee FETCHED_AT

BeagleV-Ahead

Expected Linux image path:

1. Download the latest BeagleV-Ahead distro from https://www.beagleboard.org/distros.
2. Put the board into USB flash mode by holding the USB button while connecting the USB micro-B cable, or by holding USB and pressing reset when already connected.
3. Unzip the image bundle and confirm it contains u-boot-with-spl.bin, boot.ext4, root.ext4, and fastboot_emmc.sh.
4. Flash with:

sudo ./fastboot_emmc.sh 2>&1 | tee artifacts/th1520/beaglev-ahead/logs/flash-$(date -u +%Y%m%dT%H%M%SZ).log

Expected Android status: no product Android bring-up is assumed for BeagleV-Ahead. Treat the Beagle image as a Yocto/Linux baseline unless a current official distro explicitly ships Android artifacts.

Lichee Pi 4A

Expected Linux image path:

1. Start from Sipeed's Lichee Pi 4A image summary, but prefer the current RevyOS docs when the Sipeed page redirects Linux users there.
2. Match U-Boot and DTB to board memory size: no suffix for 8 GB, 16g suffix for 16 GB.
3. Enter USB burning mode by holding BOOT while connecting USB-C to the host. On release hardware, confirm the DIP switch is in eMMC boot mode.
4. Confirm host detection:

lsusb | tee artifacts/th1520/lichee-pi-4a/logs/lsusb-burning-mode.log

5. Flash the Linux image with the image bundle's fastboot, using the image release's required fastboot when large root filesystems are involved:

sudo ./fastboot flash ram ./images/u-boot-with-spl-lpi4a-16g.bin
sudo ./fastboot reboot
sleep 1
sudo ./fastboot flash uboot ./images/u-boot-with-spl-lpi4a-16g.bin
sudo ./fastboot flash boot ./images/boot_sing.ext4
sudo ./fastboot flash root ./images/rootfs-sing.ext4

Expected Android image path:

1. Use only the Sipeed Android 13 package named by the current image summary.
2. Treat Android as experimental. A booting shell, adb devices, logcat, and surface/display proof are useful; CTS/VTS success is not expected for v0.
3. Flash the complete Android partition set:

fastboot flash ram u-boot-with-spl.bin
fastboot reboot
fastboot flash uboot u-boot-with-spl.bin
fastboot flash bootpart bootpart.ext4
fastboot flash boot boot.img
fastboot flash vendor_boot vendor_boot.img
fastboot flash super super.img
fastboot flash userdata userdata.img
fastboot flash vbmeta vbmeta.img
fastboot flash vbmeta_system vbmeta_system.img
fastboot erase metadata
fastboot erase misc

Milk-V Meles

Expected Linux image path:

1. Use Milk-V Meles official image resources for RevyOS first.
2. Select target media per test: SPI NOR firmware, microSD card image, or eMMC image.
3. For microSD, flash from the host:

xzcat <meles-image>.img.xz | sudo dd of=/dev/sdX bs=4M conv=fsync status=progress
sync

4. For eMMC, use Milk-V's eMMC installation flow and a USB-A-to-C data cable with fastboot. The board exposes Download and eMMC boot buttons; record the exact button sequence used in the log.

Expected Android status: no Android result is assumed unless Milk-V publishes a board-specific Android image. Use Meles as a Linux/RevyOS comparison board.

Serial, JTAG, And Log Capture

Serial capture is mandatory before any benchmark is valid.

mkdir -p artifacts/th1520/<board>/logs
script -f artifacts/th1520/<board>/logs/serial-$(date -u +%Y%m%dT%H%M%SZ).typescript \
  -c "tio /dev/ttyUSB0 -b 115200"

Required boot-log markers:

• ROM or SPL output, if exposed.
• U-Boot banner and DRAM size.
• OpenSBI banner.
• Linux kernel command line.
• Root filesystem mount.
• Login prompt, shell, or Android adb authorization state.
• uname -a, /proc/cpuinfo, /proc/meminfo, lsblk, ip addr, and /proc/device-tree/model.

JTAG policy:

• JTAG is a recovery/debug tool, not the primary flashing path.
• For Lichee Pi 4A, use CKLink mode and the documented pinmux step before opening the vendor debug server:

sudo memtool mw 0xfffff4a404 0
DebugServerConsole

• For BeagleV-Ahead and Meles, first prove non-invasive JTAG attach at reset halt. Do not write memory or flash over JTAG until UART and vendor flashing recovery have been proven.
• Store JTAG transcripts under artifacts/th1520/<board>/logs/jtag-*.

Android log capture, when Android boots:

adb wait-for-device
adb shell getprop | tee artifacts/th1520/<board>/logs/android-getprop.txt
adb logcat -b all -d | tee artifacts/th1520/<board>/logs/android-logcat.txt
adb shell dmesg | tee artifacts/th1520/<board>/logs/android-dmesg.txt

Linux log capture:

uname -a | tee artifacts/th1520/<board>/logs/uname.txt
cat /proc/cpuinfo | tee artifacts/th1520/<board>/logs/cpuinfo.txt
cat /proc/meminfo | tee artifacts/th1520/<board>/logs/meminfo.txt
lsblk -o NAME,SIZE,TYPE,FSTYPE,MOUNTPOINTS | tee artifacts/th1520/<board>/logs/lsblk.txt
dmesg -T | tee artifacts/th1520/<board>/logs/dmesg.txt

Benchmark Commands

Run benchmarks only after the board has a heatsink/fan installed, a stable power source, and an idle temperature log. For each benchmark, capture ambient temperature, cooling setup, power method, board revision, image SHA256, kernel, governor, and clock policy.

Preparation:

mkdir -p ~/bench && cd ~/bench
sudo sh -c 'echo performance > /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor' || true
for z in /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor; do
  sudo sh -c "echo performance > $z" || true
done

CPU smoke:

git clone https://github.com/eembc/coremark.git
cd coremark
make PORT_DIR=linux XCFLAGS="-O3 -DPERFORMANCE_RUN=1" run1.log
tee ~/bench/coremark-summary.txt < run1.log

Memory bandwidth:

git clone https://github.com/jeffhammond/STREAM.git
cd STREAM
gcc -O3 -fopenmp -DSTREAM_ARRAY_SIZE=80000000 -DNTIMES=20 stream.c -o stream
OMP_NUM_THREADS=$(nproc) ./stream | tee ~/bench/stream.txt

Memory latency:

lat_mem_rd 256M 64 2>&1 | tee ~/bench/lmbench-lat-mem-rd.txt
bw_mem 512M rd 2>&1 | tee ~/bench/lmbench-bw-mem-rd.txt
bw_mem 512M wr 2>&1 | tee ~/bench/lmbench-bw-mem-wr.txt

Storage:

fio --name=seqread --filename=~/bench/fio.test --size=1G --bs=1M --rw=read \
  --direct=1 --iodepth=16 --runtime=60 --time_based --group_reporting \
  | tee ~/bench/fio-seqread.txt
fio --name=randread --filename=~/bench/fio.test --size=1G --bs=4k --rw=randread \
  --direct=1 --iodepth=32 --runtime=60 --time_based --group_reporting \
  | tee ~/bench/fio-randread.txt

Network:

iperf3 -s
# from host:
iperf3 -c <board-ip> -P 4 -t 60 | tee artifacts/th1520/<board>/bench/iperf3-client.txt

NPU and Android-adjacent smoke:

# Linux NPU visibility
lsmod | tee ~/bench/lsmod.txt
find /dev /sys -iname '*npu*' -o -iname '*thead*' | tee ~/bench/npu-nodes.txt

# If TensorFlow Lite benchmark_model is available
benchmark_model --graph=mobilenet_v1_1.0_224_quant.tflite --num_threads=4 \
  --warmup_runs=5 --num_runs=50 | tee ~/bench/tflite-cpu.txt

# If Android boots
adb shell cmd package list packages | tee artifacts/th1520/<board>/bench/android-packages.txt
adb shell am start -W com.android.settings/.Settings | tee artifacts/th1520/<board>/bench/android-settings-start.txt

Do not report NPU TOPS from marketing material as a measured project result. Report model name, delegate/runtime, precision, latency distribution, fallback rate, CPU utilization, memory bandwidth indicators, and power.

Pass/Fail Gates

Procurement pass:

• Board model and revision are photographed and recorded.
• Accessories required for power, serial, flashing, display, and Ethernet are in hand.
• Official image source URL, image filename, fetch date, and SHA256 are stored.

Boot pass:

• Board reaches U-Boot and Linux or Android userspace from the intended media.
• Serial log includes ROM/SPL or first visible bootloader output through login.
• DRAM size, storage size, Ethernet MAC, and board model match purchased configuration or have a written explanation.

Debug pass:

• UART is stable at 115200 baud for a full cold boot.
• Recovery flashing path is proven once on the sacrificial unit.
• JTAG attach reaches a non-destructive halt/read state or is explicitly marked blocked with adapter, pinout, and voltage notes.

Linux board benchmark pass:

• CoreMark, STREAM, lmbench latency/bandwidth, fio, and iperf3 complete three consecutive runs without kernel panic, thermal shutdown, filesystem remount read-only, or SSH/UART loss.
• Cooling and power setup are unchanged during all comparable runs.
• Raw logs and command transcripts are retained.

Android exploratory pass:

• Android image boots to adb devices and produces getprop, dmesg, and logcat.
• Display or screencap evidence is captured.
• Failures in camera, GPU, NPU, CTS, VTS, or NNAPI are not project blockers for v0, but must be named before making any Android compatibility claim.

Fail conditions:

• No serial capture for a claimed boot or benchmark.
• Image filename, SHA256, or board revision missing.
• Board runs without required heatsink/fan during performance tests.
• Benchmark commands are modified between boards without recording the diff.
• Any result is compared to phone-class silicon above benchmark matrix L4.

Result Record Template

board:
revision:
serial_or_qr:
ram_emmc:
power_supply:
cooling:
image_url:
image_sha256:
kernel:
boot_media:
flash_method:
uart_adapter:
jtag_adapter:
logs:
benchmarks:
pass_fail:
waivers:

Source References

• BeagleV-Ahead quick start and flashing: https://docs.beagleboard.org/boards/beaglev/ahead/02-quick-start.html
• BeagleV-Ahead design, UART, and JTAG location: https://docs.beagleboard.org/boards/beaglev/ahead/03-design.html
• Sipeed Lichee Pi 4A overview: https://wiki.sipeed.com/hardware/en/lichee/th1520/lp4a.html
• Sipeed Lichee Pi 4A image summary: https://wiki.sipeed.com/hardware/en/lichee/th1520/lpi4a/3_images.html
• Sipeed Lichee Pi 4A flashing: https://wiki.sipeed.com/hardware/en/lichee/th1520/lpi4a/4_burn_image.html
• Sipeed Lichee Pi 4A UART and JTAG: https://wiki.sipeed.com/hardware/en/lichee/th1520/lpi4a/6_peripheral.html
• Milk-V Meles boot guide: https://milkv.io/docs/meles/getting-started/boot
• Milk-V Meles installation: https://milkv.io/docs/meles/installation
• Milk-V Meles hardware: https://milkv.io/docs/meles/hardware/meles-main-board