GPU Card Details and Application Scenarios
Details of GPU Card
The basic parameter information of the NVIDIA GPU card used in GPU server is as follows:
| Type of GPU Card | CUDA Cores | Memory Capacity | Single-Precision Floating-Point Performance | Double-Precision Floating-Point Performance | INT8 Performance | INT4 Performance | Hybrid Precision |
|---|---|---|---|---|---|---|---|
| NVIDIA Tesla T4 | 2560 | 16GB | 8.1 Tflops | -- | 130 Tops | 260 Tops | -- |
| NVIDIA Tesla V100-32G | 5120 | 32GB | 15.7 Tflops | 7.8 Tflops | -- | -- | 125 Tflops |
| NVIDIA Tesla V100-16G | 5120 | 16GB | 15.7 Tflops | 7.8 Tflops | -- | -- | 125 Tflops |
| NVIDIA Tesla P40 | 3840 | 24GB | 12 Tflops | -- | 47 Tops | -- | -- |
| NVIDIA Tesla P4 | 2560 | 8GB | 5.5 Tflops | -- | 22 Tops | -- | -- |
| NVIDIA Tesla K40 | 2880 | 12GB | 4.29 Tflops | 1.43 Tflops | -- | -- | -- |
| NVIDIA deep learning development card | 3584 | 12GB | 11 Tflops | -- | 44 Tops | -- | -- |
GPU Application Scenarios
For the off-line training scenario of deep learning, if you are a beginner of deep learning or a start-up, we recommend you the NVIDIA deep learning development card which has the highest cost performance. In creating CPU Cloud Compute, coordinating with Integrated GPU Drive Image, which can help you rapidly finish the environmental deployment and Training experiment.
For the off-line training scenario of deep learning, if you are the heavy user of the deep learning, and periodically need to train large amount of data, we recommend you the NVIDIA Tesla P40, which has better performance and stability. If you have the requirement of performance perfection, NVIDIA Tesla T4, NVIDIA Tesla V100 and the newest GPU card of NVIDIA, which support the function of Tensor Core, are more preferable. You can select to use on demand. After training, they can release the resource immediately and therefore save the cost. Simultaneously, you can rapidly accelerate the service configuration and constructions of the cloud sever of the GPU by user-defining image function.
For the on-line prediction scenario of deep learning, its requirement for the GPU performance is lower than off-line training. However it requires higher operating stability and lower response delay to server. Therefore we recommend you NVIDIA Tesla T4 and NVIDIA Tesla P4 which provide more cost-effective options while meeting performance requirements.
Of course, what is the best is not the best cloud resource but the cloud resource most appropriate for the business. Therefore we provide multiple types of GPU. We help you select different cloud resource according to your own business features and your budget of resource input. When your requirement on training ability is not that high, NVIDIA Tesla K40 is a good choice. When your requirement on off-line training is super high, you can also apply for GPU Baidu Baremetal Compute, It can provide a GPU cluster with 100 G Interconnected Network, and accelerate the training task for you.