Overview
The NVIDIA T4 was released in 2018 and is the classic product for cloud AI inference, deployed extensively from 2018 to 2022. Based on the Turing architecture, its 70W TDP supports passive cooling and virtual GPU (vGPU), making it the mainstream GPU for cloud provider inference instances. It has been superseded by the NVIDIA L4.
Core Specifications
| Item | Spec |
|---|
| Architecture | Turing TU104 |
| Process | TSMC 12nm |
| Memory | 16 GB GDDR6 |
| Memory Bandwidth | 320 GB/s |
| CUDA Cores | 2,560 |
| Tensor Cores | 320 (2nd Gen) |
| FP32 | 8.1 TFLOPS |
| INT8 Tensor Core | 130 TOPS |
| FP16 Tensor Core | 65 TFLOPS |
| TDP | 70 W (passive cooling capable) |
| PCIe | Gen 3 x16 |
| vGPU | Supported |
Use Cases
- Cloud AI inference (vGPU)
- Video transcoding
- Virtual desktop (VDI)
- Edge inference