深度学习训练,选择P100就对了

本文通过NVCaffe、MXNet、TensorFlow框架对比了NVIDIA Tesla P100和P40在深度学习训练中的性能。尽管P40在单精度运算能力上占优,但由于P100拥有更高的显存带宽,实测显示P100在深度学习训练中的性能至少高出20%,因此对于深度学习训练,P100是更好的选择。

摘要生成于 C知道 ,由 DeepSeek-R1 满血版支持, 前往体验 >

1.背景

去年4月,NVIDIA推出了Tesla P100加速卡,速度是NVIDIA之前高端系统的12倍。同年9月的GTC China 2016大会,NVIDIA又发布了Tesla P4、P40两款深度学习芯片。Tesla P100主攻学习和训练任务,而Tesla P4&P40主要负责图像、文字和语音识别。

同为Pascal架构且运算能力接近的P100和P40常常被拿来对比,单看Spec上运算能力,似乎P40比P100的深度学习性能更好,但实际上呢?本文就通过使用NVCaffe、MXNet、TensorFlow三个主流开源深度学习框架对P100和P40做性能实测来揭晓答案吧。

2.初步分析

我们先来看下P100和P40的Spec参数指标。从参数来看,的确是P40的单精运算能力强于P100,而深度学习训练普遍使用单精度浮点类型,是衡量深度学习性能的一个重要指标。P40支持IN8类型,所以非常适合精度要求不高的推理场景,24G的显存也非常适合在线处理大数据量的图像等。但是P100搭载的是HBM2高速显存,而P40只搭载了GDDR5的显存,这使得P100的显存带宽达到了P40的2倍多,

李宏毅2020机器学习深度学习 P1. Machine Learning 2020_ Course Introduction P2. Rule of ML 2020 P3. Regression - Case Study P4. Basic Concept P5. Gradient Descent_1 P6. Gradient Descent_2 P7. Gradient Descent_3 P8. Optimization for Deep Learning 1_2 选学 P9. Optimization for Deep Learning 2_2 选学 P10. Classification_1 P11. Logistic Regression P12. Brief Introduction of Deep Learning P13. Backpropagation P14. Tips for Training DNN P15. Why Deep- P16. PyTorch Tutorial P17. Convolutional Neural Network P18. Graph Neural Network 1_2 选学 P19. Graph Neural Network 2_2 选学 P20. Recurrent Neural Network Part I P21. Recurrent Neural Network Part II P22. Unsupervised Learning - Word Embedding P23. Transformer P24. Semi-supervised P25. ELMO, BERT, GPT P26. Explainable ML 1_8 P27. Explainable ML 2_8 P28. Explainable ML 3_8 P29. Explainable ML 4_8 P30. Explainable ML 5_8 P31. Explainable ML 6_8 P32. Explainable ML 7_8 P33. Explainable ML 8_8 P34. More about Explainable AI 选学 P35. Attack ML Models 1_8 P36. Attack ML Models 2_8 P37. Attack ML Models 3_8 P38. Attack ML Models 4_8 P39. Attack ML Models 5_8 P40. Attack ML Models 6_8 P41. Attack ML Models 7_8 P42. Attack ML Models 8_8 P43. More about Adversarial Attack 1_2 选学 P44. More about Adversarial Attack 2_2 选学 P45. Network Compression 1_6 P46. Network Compression 2_6 P47. Network Compression 3_6 P48. Network Compression 4_6 P49. Network Compression 5_6 P50. Network Compression 6_6 P51. Network Compression 1_2 - Knowledge Distillation .flv P52. Network Compression 2_2 - Network Pruning 选学 P53. Conditional Generation by RNN & Attention P54. Pointer Network P55. Recursive P56. Transformer and its variant 选学 P57. Unsupervised Learning - Linear Methods P58. Unsupervised Learning - Neighbor Embedding P59. Unsupervised Learning - Auto-encoder P60. Unsupervised Learning - Deep Generative Model Part.flv P61. Unsupervised Learning - Deep Generative Model Part.flv P62. More about Auto-encoder 1_4 P63. More about Auto-encoder 2_4 P64. More about Auto-encoder 3_4 P65. More about Auto-encoder 4_4 P66. Self-supervised Learning 选学 P67. Anomaly Detection 1_7 P68. Anomaly Detection 2_7 P69. Anomaly Detection 3_7 P70. Anomaly Detection 4_7 P71. Anomaly Detection 5_7 P72. Anomaly Detection 6_7 P73. Anomaly Detection 7_7 P74. More about Anomaly Detection 选学 P75. Generative Adversarial Network1_10 P76. Generative Adversarial Network2_10 P77. Generative Adversarial Network3_10 P78. Generative Adversarial Network4_10 P79. Generative Adversarial Network5_10 P80. Generative Adversarial Network6_10 P81. Generative Adversarial Network7_10 P82. Generative Adversarial Network8_10 P83. Generative Adversarial Network9_10 P84. Generative Adversarial Network10_10 P85. SAGAN, BigGAN, SinGAN, GauGAN, GANILLA, NICE-GAN(选学.flv P86. Transfer Learning P87. More about Domain Adaptation 1_2 选学 P88. More about Domain Adaptation 2_2 选学 P89. Meta Learning – MAML 1_9 P90. Meta Learning – MAML 2_9 P91. Meta Learning – MAML 3_9 P92. Meta Learning – MAML 4_9 P93. Meta Learning – MAML 5_9 P94. Meta Learning – MAML 6_9 P95. Meta Learning – MAML 7_9 P96. Meta Learning – MAML 8_9 P97. Meta Learning – MAML 9_9 P98. More about Meta Learning 选学 P99. More about Meta Learning 选学 P100. Life Long Learning 1_7 P101. Life Long Learning 2_7 P102. Life Long Learning 3_7 P103. Life Long Learning 4_7 P104. Life Long Learning 5_7 P105. Life Long Learning 6_7 P106. Life Long Learning 7_7 P107. Deep Reinforcemen Learning3_1 P108. Deep Reinforcemen Learning3_2 P109. Deep Reinforcemen Learning3_3 P110. RL Advanced Version_1_Policy Gradient P111. RL Advanced Version_2_ Proximal Policy Optimizatio.flv P112. RL Advanced Version_3_Q-Learning P113. RL Advanced Version_4_Q-Learning Advanced Tips P114. RL Advanced Version_5_Q-Learning Continuous Action.flv P115. RL Advanced Version_6_Actor-Critic P116. RL Advanced Version_7_Sparse Reward P117. RL Advanced Version_8_Imitation Learning
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值