HSIMCHUMAN–AI REALITY INFRASTRUCTURE
T736 / Complete work ecosystem / NPU acceleration SKU

C0-002 · REAL HARDWARE EVIDENCE · 2026-10-10

First NPU inference through MCP.T736 首次真实 NPU × MCP 推理验收。

A fixed vendor reference model was run on actual T736 NPU hardware via an independently installed MCP Client and an on-device authenticated-SSH stdio Server. The same work also passed a direct operator Terminal run.从离线驱动编译，到真实 NPU 推理，再到 MCP 可寻址、可执行与可回读。这是指定硬件及固定测试模型的一次通过，不是万能 AI 或整套产品交付认证。

Read verified evidence JSON / 查看证据Back to T736-ECO-002
12.827 msHardware-reported inference, MCP run / MCP 调用中的单次硬件推理
PASSFixed NBG inference / 固定模型推理执行
5MCP tools / 已验收工具
5MCP resources / 已验收资源
Exact acceptance boundary: one fixed model and one fixed input on the tested running node. No Golden Tensor was supplied. Correctness/accuracy, arbitrary models, sustained throughput, reboot persistence and general work authority are not accepted.仅通过固定样例的真实执行；没有 Golden Tensor 对照，尚不宣称结果正确率、任意模型兼容、重启持久性或生产级开放服务。

Actual execution chain已经实际通过的计算与协议层次

  1. 01 · HARDWARE

    Allwinner T736真实 NPU 节点

    Device ID 0x1000003B; 1 NPU core.

  2. 02 · KERNEL

    vipcore driver驱动绑定与设备节点

    Linux 5.15.119, /dev/vipcore. Manual load, not persistent.

  3. 03 · RUNTIME

    VIPLite v2用户态运行库

    2.0.3.2-AW-2024-08-30; isolated working directory.

  4. 04 · MODEL

    Fixed NBG经匹配的原始测试模型

    Model target CID 0x1000003B; single run.

  5. 05 · MCP

    npu.infer.fixed经认证的模型执行工具

    Modern MCP 2026-07-28 over SSH stdio.

  6. 06 · EVIDENCE

    Readback操作证据返回

    Operation ID, output log hash, timings and latest resource.

Two independent runs两次真实 NPU 推理的明确区分

RunProfile timeInterpretation
Operator Terminal / 真人触发12.831 msFixed NBG executed; vpm run ret=0; golden 0
Official MCP Client / MCP 触发12.827 msOn-device tool completed, resource readback matched operation ID

Times are per-run reports under this exact input; not a statistically valid benchmark or performance SLA.

Why this changes the work relationship改变的不是芯片，而是能力如何被接续

Previously, a developer had to find the correct BSP, enable the kernel driver, locate a compatible userspace library and find a NBG compiled for the correct NPU CID. These exact dependencies were established from source material and tested on the actual hardware. The result is now described as one addressable, bounded MCP computation instead of an undocumented Terminal-only experiment.

e-Person 能够通过受限、可核对的能力入口使用本地计算；但它不能因为这次成功自动获得摄像头、机电执行、全域知识或跨地点授权。每一种能力仍要有明确用途、边界与回读证据。

Continue from the complete-work ecosystem →

Follow-up evidence: Golden + reboot recovery后续已验收：Golden 与重启恢复

The earlier fixed-model MCP execution has a new, separately scoped follow-up: one reboot, automatic vipcore loading, post-boot MCP runs and an independent 256-byte vendor Golden comparison on a different custom_lut test model.

Open accepted C0-003 evidence →

What comes next尚待形成的能力与证据

Verify a Golden Tensor or task-specific output; exercise reboot and recovery; add model compatibility checks and sustained thermal/power tests; decide how authenticated users and work roles receive scoped capability access; optionally validate a secured network transport. This first MCP tool does not expose a public endpoint or accept arbitrary model paths and inputs.

No file from the proprietary vendor BSP, model binary, personal credentials or private runtime log is published on this page. Evidence is summarized with hashes and bounded scope.