Neural Compute Stick 2性能测试

简介

对二代计算棒，官方宣布其性能是一代的8倍。

测试结果

测试代码使用的是官方提供的benchmark_app
Count：执行的次数
Duration：执行Count总共使用的时间
Latency：从该次数据输入到得到该结果的时间
Throughput：就是在同步，或者异步条件下所能达到的实际速度。
inference requests: 我目前认为是计算棒能加载几个网络模型，二代是能加载4个的。测试时有意进行了限制为2的测试，方便和一代对比。

USB2.0
Full device name: Intel Movidius Myriad 2 VPU
[Step 10/11] Measuring performance (Start inference asyncronously, 2 inference requests, limits: 1000 iterations)
Count:      1000 iterations
Duration:   25783.63 ms
Latency:    51.54 ms
Throughput: 38.78 FPS

[Step 10/11] Measuring performance (Start inference syncronously, limits: 1000 iterations)

[Step 11/11] Dumping statistics report
Count:      1000 iterations
Duration:   30224.62 ms
Latency:    30.22 ms
Throughput: 33.09 FPS

Full device name: Intel Movidius Myriad X VPU
[Step 10/11] Measuring performance (Start inference asyncronously, 2 inference requests, limits: 1000 iterations)

[Step 11/11] Dumping statistics report
Count:      1000 iterations
Duration:   6170.91 ms
Latency:    12.31 ms
Throughput: 162.05 FPS
[Step 10/11] Measuring performance (Start inference asyncronously, 4 inference requests, limits: 1000 iterations)

[Step 11/11] Dumping statistics report
Count:      1000 iterations
Duration:   5580.84 ms
Latency:    22.27 ms
Throughput: 179.18 FPS

[Step 10/11] Measuring performance (Start inference syncronously, limits: 1000 iterations)

[Step 11/11] Dumping statistics report
Count:      1000 iterations
Duration:   11834.24 ms
Latency:    11.83 ms
Throughput: 84.51 FPS


USB3.0
Full device name: Intel Movidius Myriad 2 VPU
[Step 10/11] Measuring performance (Start inference asyncronously, 2 inference requests, limits: 1000 iterations)

[Step 11/11] Dumping statistics report
Count:      1000 iterations
Duration:   25867.01 ms
Latency:    51.71 ms
Throughput: 38.66 FPS

[Step 10/11] Measuring performance (Start inference syncronously, limits: 1000 iterations)

[Step 11/11] Dumping statistics report
Count:      1000 iterations
Duration:   27204.58 ms
Latency:    27.20 ms
Throughput: 36.77 FPS

Full device name: Intel Movidius Myriad X VPU
[Step 10/11] Measuring performance (Start inference asyncronously, 2 inference requests, limits: 1000 iterations)

[Step 11/11] Dumping statistics report
Count:      1000 iterations
Duration:   4098.44 ms
Latency:    8.19 ms
Throughput: 244.00 FPS

[Step 10/11] Measuring performance (Start inference asyncronously, 4 inference requests, limits: 1000 iterations)

[Step 11/11] Dumping statistics report
Count:      1000 iterations
Duration:   3483.48 ms
Latency:    13.91 ms
Throughput: 287.07 FPS

[Step 10/11] Measuring performance (Start inference syncronously, limits: 1000 iterations)

[Step 11/11] Dumping statistics report
Count:      1000 iterations
Duration:   7649.50 ms
Latency:    7.63 ms
Throughput: 130.99 FPS

结论

在USB3.0条件下，二代计算棒大约是一代的7.5倍速度。和官宣相差不多。