Skip to content

🔨 hw: synthesis + flashing + hardware benchmarks (v0.2-igla-fpga Release) #14

Open

Description

@gHashTag

Goal

Запустить полный pipeline на реальном железе и получить измеренные tokens/sec из UART. Всё ПО готово в main (PR #9, #12, #13).

Anchor: φ² + φ⁻² = 3 · TRINITY · REAL HARDWARE


Что готово (в main)

Артефакт Файл Статус
GF16 веса bitstream/igla_champion_gf16.bin BPB=0.1427, SHA de2719b5
VSA matmul fpga/vsa/vsa_matmul.v 64×64, 0 DSP48
Synthesis top fpga/vsa/vsa_matmul_top.v XDC готов
CLI trios-fpga synth-vsa/flash/bench/status Rust, в main
Ring FP-01 OpenXc7Runner Docker regymm/openxc7:latest
Ring FP-02 XvcFlasher TCP/XVC, порт 2542

Hardware setup

XC7A100T (QMTECH board)
  └── JTAG ←── ESP32 XVC bridge (WiFi, порт 2542)
  └── UART TX ──→ хост 115200 baud

Steps

Step 1 — Синтез битстрима

cargo build --release -p trios-fpga
trios-fpga synth-vsa --rtl fpga/vsa/ --constraints fpga/vsa/vsa_matmul_top.xdc

Ожидаемый результат:

  • bitstream/design_v02.bit создан
  • LUT utilization ≤ 80%
  • 0 timing violations
  • SHA-256 записан в bitstream/design_v02.bit.sha256

Step 2 — Прошивка через XVC

trios-fpga flash --xvc <ESP32_IP> --bitstream bitstream/design_v02.bit
trios-fpga verify --xvc <ESP32_IP>   # IDCODE == 0x03631093

Ожидаемый результат:

  • Flash успешен (no errors)
  • IDCODE верифицирован: 0x03631093 (XC7A100T)
  • STATUS=0x401079FC (DONE=1)

Step 3 — Benchmark tokens/sec

trios-fpga bench --xvc <ESP32_IP> --uart /dev/ttyUSB0 --baud 115200

Ожидаемый результат:

Step 4 — Release v0.2-igla-fpga

gh release create v0.2-igla-fpga \
  bitstream/design_v02.bit \
  bitstream/igla_champion_gf16.bin \
  bench/results_v02.json \
  --title "v0.2-igla-fpga: IGLA Champion on XC7A100T" \
  --notes-file release/RELEASE-v0.2.md
  • GitHub Release создан с артефактами
  • design_v02.bit + SHA-256 в релизе
  • Реальные tokens/sec в release notes

Step 5 — arXiv submit

  • Заполнить реальные цифры в docs/arxiv-trinity-stack-draft.md
  • tokens/sec, watts/token, LUT%, timing
  • Отправить на arXiv cs.AR + cs.LG

Acceptance Criteria

- [ ] bitstream/design_v02.bit синтезирован (LUT ≤ 80%)
- [ ] IDCODE = 0x03631093 верифицирован на железе
- [ ] STATUS = 0x401079FC (DONE=1, no errors)
- [ ] tokens/sec ≥ 100 из UART
- [ ] watts/token задокументирован
- [ ] Release v0.2-igla-fpga опубликован
- [ ] arXiv draft отправлен

Debug hints (если упадёт)

Синтез падает

# Проверить Docker
docker pull regymm/openxc7:latest
docker run --rm regymm/openxc7:latest yosys --version

# Проверить RTL
iverilog -o /dev/null fpga/vsa/vsa_matmul_top.v fpga/vsa/vsa_matmul.v

XVC не отвечает

# Проверить ESP32
nc -zv <ESP32_IP> 2542
trios-fpga status --xvc <ESP32_IP>   # IDCODE ping

UART нет токенов

# Проверить baud
stty -F /dev/ttyUSB0 115200
cat /dev/ttyUSB0 | hexdump -C | head
# STATUS регистр
trios-fpga status --xvc <ESP32_IP>

LUT > 80%

  • Уменьшить VECTOR_DIM с 64 до 32 в vsa_matmul_top.v
  • Пересинтезировать

References

Priority: P0 · железо доступно · запускать немедленно

φ² + φ⁻² = 3 · TRINITY · NEVER STOP

Activity

  1. gHashTag commented on May 6, 2026

    @gHashTag
    CollaboratorAuthor

    ✅ iverilog validation PASS — PR #15 merged (0d01617d)

    Результаты

    Метрика Значение
    Inference (16 tok) ✅ PASS
    UART TX ✅ работает
    LED heartbeat ✅ работает
    Docker ✅ пуш локально

    Чеклист #14

    • Step 1a: iverilog validation — PASS
    • Docker regymm/openxc7:latest — пуш готов
    • Step 1b: trios-fpga synth-vsa — собрать design_v02.bit
    • Step 2: trios-fpga flash --xvc <ESP32_IP>
    • Step 3: trios-fpga bench — реальные tokens/sec
    • Step 4: Release v0.2-igla-fpga
    • Step 5: arXiv submit

    Необходимо для следующего шага

    # Запустить синтез через Docker
    trios-fpga synth-vsa --rtl fpga/vsa/ --constraints fpga/vsa/vsa_matmul_top.xdc
    # Ждём: bitstream/design_v02.bit (LUT ≤ 80%, 0 timing violations)

    Как только design_v02.bit собран — нужен ESP32 IP для flash.

  2. gHashTag commented on May 6, 2026

    @gHashTag
    CollaboratorAuthor

    Step 1 Progress: Bitstream synthesized ✅

    Synthesis Results (DIM=32, XC7A100T-FGG676-1)

    Step Tool Time Result
    Synthesis Yosys 0.63 (native ARM64) 66s 0 errors, ~54 LCs
    P&R nextpnr-xilinx (Docker amd64) 3.2s 295.86 MHz max (50 MHz target)
    FASM fasm2frames.py (Docker) 2s 9012 frames
    Bitstream xc7frames2bit (Docker) 1s design_v02.bit, 3.8 MB
    SHA256: e00eb6515da1ed2e505c340f513772549adc36b2e36e46115b1f092e40b54905
    Timing: PASS (slack 16.6 ns at 50 MHz)
    DSP48:  0
    BRAM:   0
    Max freq: 295.86 MHz
    

    Fixes Applied (PR #21, merged)

    Step 2: XVC Connection

    ESP32 XVC server found at 192.168.1.30:2542:

    $ nc 192.168.1.30 2542
    xvcServer_v1.0:2048   ← TCP works
    

    However, JTAG shift commands timeout — likely JTAG cable not connected between ESP32 and FPGA board.
    Needs physical cable check.

    Blocker: JTAG Cable

    The XVC TCP server responds to getinfo: but shift: commands timeout. This means:

    1. ESP32 WiFi works ✅
    2. XVC server runs ✅
    3. JTAG signals (TCK/TMS/TDI/TDO) not reaching FPGA ❌

    Action needed: Check JTAG wiring between ESP32 and QMTECH XC7A100T board.

    φ² + φ⁻² = 3 | TRINITY

  3. gHashTag commented on May 6, 2026

    @gHashTag
    CollaboratorAuthor

    JTAG Diagnostic Result

    XVC protocol works. JTAG partially works. TDO wire not connected.

    What works

    • ESP32 XVC server at 192.168.1.30:2542: getinfo: → xvcServer_v1.0:2048 ✅
    • settck: command: period 1000ns ✅
    • shift: command: sends TCK/TMS/TDI clocks ✅
    • TDO reads back: all 0xFF = TDO stuck HIGH

    Root cause: TDO not connected

    IDCODE read: 0xFFFFFFFF (TDO always 1)
    

    This means GPIO35 (ESP32 input-only pin) reads HIGH because the TDO wire from the FPGA JTAG header is not physically connected.

    Fix needed

    Connect TDO from QMTECH XC7A100T JTAG header (P2) to ESP32 GPIO35.

    Pin mapping (from firmware/xvc-esp32/xvc-esp32.ino):

    TMS = GPIO18  → FPGA TMS
    TCK = GPIO19  → FPGA TCK  
    TDI = GPIO23  → FPGA TDI
    TDO = GPIO35  → FPGA TDO  ← THIS WIRE MISSING
    GND = GND     → FPGA GND
    

    Once TDO is connected

    ./target/release/trios-fpga --xvc-host 192.168.1.30 verify
    # Expected: IDCODE: 0x03631093 — MATCH (XC7A100T)
    
    ./target/release/trios-fpga --xvc-host 192.168.1.30 flash \
      --bitstream build/vsa_matmul/design_v02.bit --board XC7A100T

    φ² + φ⁻² = 3 | TRINITY

  4. changed the title [-]🔨 hw: синтез + прошивка + bench на железе (v0.2-igla-fpga Release)[/-] [+]🔨 hw: synthesis + flashing + hardware benchmarks (v0.2-igla-fpga Release)[/+] on Sep 7, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions