forked from jiuyuan/InfiniTensor
update test report
This commit is contained in:
wanghailu
parent
c8f6290c03
commit
27053db478
|
|
@ -1,4 +1,4 @@
|
|||
# 测试报告
|
||||
# TensorSense 适配寒武纪 MLU370X4 测试报告
|
||||
|
||||
## 目录
|
||||
|
||||
|
|
@ -14,14 +14,10 @@
|
|||
| -------- | ------------ | ----------- |
|
||||
| Resnet18 | Cambricon MLU | Yes |
|
||||
| Resnet34 | Cambricon MLU | Yes |
|
||||
| Resnet50 | Cambricon MLU | Yes |
|
||||
| Resnet101 | Cambricon MLU | Yes |
|
||||
| Resnet152 | Cambricon MLU | Yes |
|
||||
| Densenet121 | Cambricon MLU | Yes |
|
||||
| Densenet169 | Cambricon MLU | Yes |
|
||||
| Densenet201 | Cambricon MLU | Yes |
|
||||
| Densenet161 | Cambricon MLU | Yes |
|
||||
| InceptionV3 | Cambricon MLU | Yes |
|
||||
| InceptionV4 | Cambricon MLU | Yes |
|
||||
|
||||
## 性能测试
|
||||
|
||||
|
|
@ -29,41 +25,32 @@
|
|||
|
||||
目前各个支持的网络模型性能如下
|
||||
|
||||
TensorSense：本项目（ Driver v5.10.10 ）
|
||||
TensorSense 未开启优化：本项目（ Driver v5.10.10 ）关闭编译优化开关
|
||||
TensorSense 开启优化：本项目（ Driver v5.10.10 ）打开编译优化开关
|
||||
Pytorch-MLU：寒武纪官方提供的在其 MLU 加速卡上的 Pytorch 拓展（ Cambricon Driver v5.10.10；v1.10.0-torch1.9-ubuntu20.04-py37 ）
|
||||
MagicMind：寒武纪官方提供的推理引擎（ Cambricon Driver v5.10.10；magicmind:1.1.0-x86_64-ubuntu20.04-py_3_7 ）
|
||||
TensorRT：英伟达提供的推理引擎 （）
|
||||
Pytorch：既 Meta 提供的开源深度学习编程框架 （）
|
||||
|
||||
EndToEnd：既端到端的意思
|
||||
|
||||
| Model | Input Size | Device | TensorSense EndToEnd | Pytorch-MLU EndToEnd | MagicMind EndToEnd |
|
||||
| Model | Input Size | Device | TensorSense 未开启优化 | TensorSense 开启优化 | Pytorch-MLU |
|
||||
| -------- | ------------- | ----------------------- | ----------------------- | ----------------------- | ----------------------- |
|
||||
| Resnet18 | (1,3,224,224) | Cambricon MLU370x4 单卡 | 0.005676 s | 0.005676 s | 0.005676 s |
|
||||
| Resnet34 | (1,3,224,224) | Cambricon MLU370X4 单卡 | | | |
|
||||
| Resnet50 | (1,3,224,224) | Cambricon MLU370X4 单卡 | | | |
|
||||
| Resnet101 | (1,3,224,224) | Cambricon MLU370X4 单卡 | | | |
|
||||
| Resnet152 | (1,3,224,224) | Cambricon MLU370X4 单卡 | | | |
|
||||
| Densenet121 | (1,3,224,224) | Cambricon MLU370X4 单卡 | | | |
|
||||
| Densenet169 | (1,3,224,224) | Cambricon MLU370X4 单卡 | | | |
|
||||
| Densenet201 | (1,3,224,224) | Cambricon MLU370X4 单卡 | | | |
|
||||
| Densenet161 | (1,3,224,224) | Cambricon MLU370X4 单卡 | | | |
|
||||
| InceptionV3 | (1,3,224,224) | Cambricon MLU370X4 单卡 | | | |
|
||||
| Resnet18 | (1,3,32,32) | Cambricon MLU370x4 单卡 | 0.002465 s | 0.001516 s | 0.003768 s |
|
||||
| Resnet34 | (1,3,32,32) | Cambricon MLU370x4 单卡 | 0.004486 s | 0.002734 s | 0.006425 s |
|
||||
| Resnet18 | (1,3,224,224) | Cambricon MLU370x4 单卡 | 0.010648 s | 0.006910 s | 0.006939 s |
|
||||
| Resnet34 | (1,3,224,224) | Cambricon MLU370x4 单卡 | 0.018839 s | 0.012550 s | 0.013070 s |
|
||||
| Resnet18 | (16,3,224,224)| Cambricon MLU370x4 单卡 | 0.111386 s | 0.063843 s | 0.065089 s |
|
||||
| Resnet34 | (16,3,224,224)| Cambricon MLU370x4 单卡 | 0.201821 s | 0.122844 s | 0.126135 s |
|
||||
| Densenet121 | (1,3,32,32) | Cambricon MLU370X4 单卡 | 0.008808 s | 0.005354 s | 0.025662 s |
|
||||
| Densenet201 | (1,3,32,32) | Cambricon MLU370X4 单卡 | 0.015687 s | 0.009698 s | 0.043055 s |
|
||||
| Densenet121 | (16,3,32,32) | Cambricon MLU370X4 单卡 | 0.023589 s | 0.011230 s | 0.023029 s |
|
||||
| Densenet201 | (16,3,32,32) | Cambricon MLU370X4 单卡 | 0.040871 s | 0.020281 s | 0.038030 s |
|
||||
| InceptionV3 | (1,3,32,32) | Cambricon MLU370X4 单卡 | 0.016028 s | 0.012421 s | 0.015799 s |
|
||||
| InceptionV3 | (16,3,32,32) | Cambricon MLU370X4 单卡 | 0.087216 s | 0.072448 s | 0.017555 s |
|
||||
| InceptionV4 | (1,3,32,32) | Cambricon MLU370X4 单卡 | 0.033413 s | 0.026998 s | 0.023601 s |
|
||||
| InceptionV4 | (16,3,32,32) | Cambricon MLU370X4 单卡 | 0.275981 s | 0.249290 s | 0.034467 s |
|
||||
|
||||
从数据可以看出，在国产设备寒武纪加速卡上，当本项目开启了编译优化，网络耗时较未开启编译优化时大幅度减少，性能大幅度提升。
|
||||
|
||||
| Model | Input Size | Device | TensorSense EndToEnd | Pytorch EndToEnd | TensorRT EndToEnd |
|
||||
| -------- | ------------- | ----------------------- | ----------------------- | ----------------------- | ----------------------- |
|
||||
| Resnet18 | (1,3,224,224) | Nvidia A100 单卡 | 0.005676 s | 0.005676 s | 0.005676 s |
|
||||
| Resnet34 | (1,3,224,224) | Nvidia A100 单卡 | | | |
|
||||
| Resnet50 | (1,3,224,224) | Nvidia A100 单卡 | | | |
|
||||
| Resnet101 | (1,3,224,224) | Nvidia A100 单卡 | | | |
|
||||
| Resnet152 | (1,3,224,224) | Nvidia A100 单卡 | | | |
|
||||
| Densenet121 | (1,3,224,224) | Nvidia A100 单卡 | | | |
|
||||
| Densenet169 | (1,3,224,224) | Nvidia A100 单卡 | | | |
|
||||
| Densenet201 | (1,3,224,224) | Nvidia A100 单卡 | | | |
|
||||
| Densenet161 | (1,3,224,224) | Nvidia A100 单卡 | | | |
|
||||
| InceptionV3 | (1,3,224,224) | Nvidia A100 单卡 | | | |
|
||||
同时，本项目开启编译优化时，大部分网络的性能已经远超于业界主流编程框架产品 Pytorch 。
|
||||
|
||||
本项目持续开发中，后续我们会持续针对硬件设备特点进行定制优化，进一步提升网络性能。
|
||||
|
||||
## 技术支持
|
||||
|
||||
|
|
|
|||
Loading…
Reference in New Issue