RK3588+RK1828多模态AI问答机器人性能实测

测试代码：

#!/usr/bin/env python3

import json, time, subprocess, urllib.request, os

MODELS = ["Qwen2.5-3B", "Qwen3-VL-2B", "Qwen2.5-VL-7B", "Qwen3-8B"]

BASE = "http://127.0.0.1:8080"

PROMPT = "请用大约80个汉字介绍RK3588芯片的主要特点和典型应用场景。"

MAX_TOKENS = 128

RUNS = 2

def http_json(url, data=None, timeout=300):    

req = urllib.request.Request(url, method="GET" if data is None else "POST")    

if data is not None:

     body = json.dumps(data).encode()

     req.add_header("Content-Type", "application/json")

     req.data = body    

with urllib.request.urlopen(req, timeout=timeout) as r:

     return json.loads(r.read().decode())

def wait_model(timeout=240):

 t0 = time.time()

 while time.time() - t0 < timeout:

     try:

         j = http_json(BASE + "/v1/models", timeout=3)

         data = j.get("data") or []

        if data:

             return data[0].get("id", ""), time.time() - t0

     except Exception:

         pass

     time.sleep(2)

 return None, time.time() - t0

def switch_model(name):

script = "/userdata/models/%s/switch_to.sh" % name

 p = subprocess.run(

     "echo 123456 | sudo -S bash %s" % script,

    shell=True, stdout=subprocess.PIPE, stderr=subprocess.STDOUT,

     text=True, timeout=120)

 return p.returncode, (p.stdout or "")[-400:]

def chat(max_tokens=MAX_TOKENS):

 t0 = time.time()

 j = http_json(BASE + "/v1/chat/completions", {

     "messages": [{"role": "user", "content": PROMPT}],

    "max_tokens": max_tokens,

     "temperature": 0.1,

     "stream": False,

 }, timeout=300)

 wall = time.time() - t0

 usage = j.get("usage") or {}

 timings = j.get("timings") or {}

 content = ((j.get("choices") or [{}])[0].get("message") or {}).get("content") or ""

 return {

     "wall_s": round(wall, 3),

     "prompt_tokens": usage.get("prompt_tokens"),

     "completion_tokens": usage.get("completion_tokens"),

     "prompt_tok_s": round(timings.get("prompt_per_second") or 0, 2),

    "gen_tok_s": round(timings.get("predicted_per_second") or 0, 2),

     "prompt_ms": round(timings.get("prompt_ms") or 0, 1),

    "gen_ms": round(timings.get("predicted_ms") or 0, 1),

    "reply_preview": content.replace("\n", " ")[:100],

     "model_id": j.get("model"),

}

def model_size_gb(name):

 p = "/userdata/models/%s" % name

 out = subprocess.check_output(["du", "-sbL", p], text=True).split()[0]

 return round(int(out) / (1024 ** 3), 2)

results = []

for name in MODELS:

 print("\n===== %s =====" % name, flush=True)

 row = {"name": name, "size_gb": model_size_gb(name), "type": "VLM" if "VL" in name else "LLM"}

 t_sw = time.time()

 rc, out = switch_model(name)

 mid, load_wait = wait_model(240)

 row["load_s"] = round(time.time() - t_sw, 1)

 row["load_wait_s"] = round(load_wait, 1)

 row["model_id"] = mid

 row["switch_rc"] = rc

 print("loaded id=%s load_s=%.1f" % (mid, row["load_s"]), flush=True)

 if not mid:

    row["error"] = "load timeout"

     results.append(row)

     continue

 try:

     chat(32)

 except Exception as e:

     row["error"] = "warmup failed: %s" % e

     results.append(row)

     continue    runs = []

 for i in range(RUNS):

     try:

         r = chat(MAX_TOKENS)

         runs.append(r)

         print("  run%d: gen=%.2f tok/s prompt=%.2f tok/s wall=%.3fs n=%s" % (

             i + 1, r["gen_tok_s"], r["prompt_tok_s"], r["wall_s"], r["completion_tokens"]), flush=True)        except Exception as e:

         print("  run%d err: %s" % (i + 1, e), flush=True)

 if runs:

     row["gen_tok_s_avg"] = round(sum(r["gen_tok_s"] for r in runs) / len(runs), 2)

     row["prompt_tok_s_avg"] = round(sum(r["prompt_tok_s"] for r in runs) / len(runs), 2)

     row["wall_s_avg"] = round(sum(r["wall_s"] for r in runs) / len(runs), 3)

     row["completion_tokens_avg"] = round(sum((r["completion_tokens"] or 0) for r in runs) / len(runs), 1)

     row["prompt_tokens"] = runs[0]["prompt_tokens"]

     row["reply_preview"] = runs[-1]["reply_preview"]

     row["runs"] = runs    results.append(row)print("\n===== SUMMARY_JSON =====")print(json.dumps(results, ensure_ascii=False, indent=2))

浏览器 (局域网 :5000)

 │  SSE /query、上传图、ASR、设置

 ▼

Flask app.py  ──openai──►  rkllm3-server :8080  ──►  RK182x NPU

 │                                          ▲

 │                         │ switch_to.sh 改软链后 systemctl restart

 │                         │    

 ├─ TTS 队列 (edge-tts → espeak 回退 → ffplay)    

 └─ Unix DGRAM /tmp/dh_ipc.sock ──► digital_man.service（可选）

两套 systemd 服务：

rkllm3.service

依赖 rknn3.service，跑 /userdata/models/rkllm3_server.sh → 当前模型的

run.sh → rkllm3-server

web.service

python3 /userdata/web/app.py --webbrowser，配 PulseAudio/TTS 环境，root 跑

对以上RK3588+RK182X算力卡开发智能问答机器人感兴趣的，欢迎给我们致电或在线留言，另外我们还有多款算力板卡可供挑选，涵盖高中低算力，欢迎进一步咨询了解。