三个免费的GPU资源平台
注意:对于中国用户,以下介绍的网站可能需要非中国手机号、邮箱注册,请自行搜索解决。
Google Colab
Google Colab是由Google提供的在线计算平台,能够与Google Drive无缝连接,并且免费方案中可以不限时长地使用CPU以及有限制地使用GPU。用户还可以通过购买不同的付费方案来获得更高的计算资源、更长的运行时间和更多的功能。各种方案比较的表格如下(参考了这篇的介绍):
Colab Free | Colab Pro | Colab Pro+ | |
---|---|---|---|
CPU | 2 vCPU | 8 vCPU (High-RAM) 2 vCPU (Standard) |
8 vCPU (High-RAM) 2 vCPU (Standard) |
RAM | 13 GB | 53 GB (High-RAM) 13 GB (Standard) |
53 GB (High-RAM) 13 GB (Standard) |
GPU | NVIDIA T4 (15 GB) | NVIDIA T4 (15 GB) (1.84 Units/hr) NVIDIA V100 (16 GB) (4.91 Units/hr) NVIDIA L4 (22.5 GB) (4.82 Units/hr) NVIDIA A100 (40 GB) (11.77 Units/hr) |
NVIDIA T4 (15 GB) (1.84 Units/hr) NVIDIA V100 (16 GB) (4.91 Units/hr) NVIDIA L4 (22.5 GB) (4.82 Units/hr) NVIDIA A100 (40 GB) (11.77 Units/hr) |
Run in the Background | No | No | Yes |
Session Limit | 12 hr | 24 hr | 24 hr |
Total Time Limit | Unlimited (Conditional) | 100 Units/mth | 500 Units/mth |
Terminal | No | Yes | Yes |
Storage | 107 GB (without GPU) 78 GB (with GPU) At least 15 GB for Input/Output (Google Drive) |
225 GB (without GPU) 166 GB (with GPU) At least 15 GB for Input/Output (Google Drive) |
225 GB (without GPU) 166 GB (with GPU) At least 15 GB for Input/Output (Google Drive) |
Price | Free | US$10.49/month | US$52.49/month |
Rating | 3.5 |
Colab Pro方案的性价比相对较低且有一定的限制,因此更推荐使用免费方案进行练习。Colab Pro/Pro+ 更适合短时间需要大量计算的用户,如果需要长时间高性能计算,推荐选择Lambda等云GPU出租平台。
Kaggle Notebook
Kaggle Notebook是由Google旗下的计算竞赛平台Kaggle提供的服务,主要用于竞赛和数据科学。虽然属于Google旗下产品,但并未提供付费方案。Kaggle Notebook提供不限时长的CPU计算资源和每周至少30小时的GPU运行时间(每周六重置),以及200 GB私有数据集空间(单个数据集最大200 GB,公开数据集存储不受限)。具体参数如下:
Kaggle Notebook | |
---|---|
CPU | 4 vCPU |
RAM | 30 GB (without GPU) 29 GB (with GPU) |
GPU | NVIDIA T4*2 (15 GB*2) NVIDIA P100 (16 GB) |
Run in the Background | Yes |
Session Limit | 9 hr |
Total Time Limit | Unlimited (Conditional) (without GPU) At least 30 hr/wk (reset every Saturday) |
Terminal | Yes |
Storage | 57.6 GB (19.5 GB for Output) |
Price | Free |
Rating | 5 |
对于仅需免费资源的用户而言,每周30小时的运行时长足以满足大部分的需求,Kaggle无疑是最好的选择
SageMaker Studio Lab
SageMaker Studio Lab是Amazon AWS推出的免费的Jupyter Notebook服务,与AWS主账户独立,需要额外申请注册,审核通常需要1-3天。SageMaker Stuido Lab提供类似于本地Jupyter Notebook的界面和完整的功能,包含Anaconda和终端机支持。SageMaker Studio Lab仅提供免费使用,每24小时可免费使用8小时CPU和4小时GPU(运行GPU不统计CPU时长),但是GPU常会出现不可用的情况。具体参数如下:
SageMaker Studio Lab | |
---|---|
CPU | 4 vCPU |
RAM | 16 GB |
GPU | NVIDIA T4 (15 GB) |
Run in the Background | Yes |
Session Limit | 4 hr |
Total Time Limit | 8 hr/d (without GPU) 4 hr/d (with GPU) |
Terminal | Yes |
Storage | 37 GB (15 GB for Input/Output) |
Price | Free |
Rating | 4 |
虽然SageMaker Stuido Lab的配置相对较低,但是由于有完整的Jupyter Notebook环境和扩展性,同时也方便迁移到AWS上,因此方便日后进行大项目的开发,适合较为进阶的人士使用。
免费方案比较
Colab Free | Kaggle Notebook | SageMaker Studio Lab | |
---|---|---|---|
CPU | 2 vCPU | 4 vCPU | 4 vCPU |
RAM | 13 GB | 30 GB (without GPU) 29 GB (with GPU) |
16 GB |
GPU | NVIDIA T4 (15 GB) | NVIDIA T4*2 (15 GB*2) NVIDIA P100 (16 GB) |
NVIDIA T4 (15 GB) |
Run in the Background | No | Yes | Yes |
Session Limit | 12 hr | 9 hr | 4 hr |
Total Time Limit | Unlimited (Conditional) | Unlimited (Conditional) (without GPU) At least 30 hr/wk (reset every Saturday) |
8 hr/d (without GPU) 4 hr/d (with GPU) |
Terminal | No | Yes | Yes |
Storage | 107 GB (without GPU) 78 GB (with GPU) At least 15 GB for output (Google Drive) |
57.6 GB (19.5 GB for Output) | 37 GB (15 GB for Input/Output) |
Price | Free | Free | Free |
Rating | 3.5 | 5 | 4 |