NVIDIA Tesla T4 16GB Data-Center GPU
The NVIDIA Tesla T4 is a 70 W single-slot, low-profile data-center GPU built on the Turing TU104 architecture, pairing 2,560 CUDA cores and 320 Tensor Cores with 16 GB of GDDR6 on a 256-bit bus (320 GB/s).
Engineered for inference and light training at rack scale, the T4 delivers roughly 8.1 TFLOPS FP32, 65 TFLOPS FP16 (Tensor), 130 TOPS INT8 and 260 TOPS INT4, plus dedicated NVENC/NVDEC blocks for hardware video transcode. It draws its entire 70 W budget from the PCIe 3.0 ×16 slot — there is no supplemental power connector — and is passively cooled, relying on server chassis airflow.
The single-slot, low-profile form factor lets multiple T4s pack into a 1U/2U server for dense inference farms, NVIDIA vGPU virtualization and VDI. Onboard ECC on the GDDR6 protects model and activation data in production.
Delivered seller-refurbished: vBIOS and device ID verified, the GPU and GDDR6 exercised under sustained thermal load with ECC error checking, and slot power confirmed — under a 365-day owned warranty with seller-paid RMA freight and zero restocking fee.
| Product | NVIDIA Tesla T4 (board 900-2G183-0000-001) |
|---|---|
| Architecture | NVIDIA Turing (TU104) |
| CUDA cores | 2,560 |
| Tensor cores | 320 |
| Memory | 16 GB GDDR6 (with ECC) |
| Memory bus / bandwidth | 256-bit · 320 GB/s |
| FP32 | 8.1 TFLOPS |
| FP16 (Tensor) | 65 TFLOPS |
| INT8 / INT4 | 130 TOPS / 260 TOPS |
| Interface | PCIe 3.0 ×16 |
| Power (TDP) | 70 W (slot-powered, no aux connector) |
| Form factor | Single-slot, low-profile |
| Cooling | Passive (server airflow required) |
| Video engines | 1× NVENC / 2× NVDEC |
| Virtualization | NVIDIA vGPU / GRID supported |
| Condition (default) | Seller Refurbished |
| Warranty | 365-day owned warranty |
Confirm riser, fan-kit and BIOS support with the Compatibility Tool, or browse Server Components and Servers & Systems.
Certified host servers
Enterprise servers with a GPU-enabling PCIe riser and high-airflow (GPU) fan kit — HPE ProLiant DL380 Gen10, Dell PowerEdge R740xd, and Supermicro GPU nodes. The T4 is validated on HPE, Dell, Lenovo, Cisco and Supermicro NVIDIA-qualified platforms.
Airflow requirement
This is a passively cooled card: it must be installed in a server that provides directed front-to-back airflow. It is not suitable for a desktop/workstation without an approved cooling shroud.
Power
Slot-powered at 70 W — no 6/8-pin PCIe power cable required.
Cross-reference / family
900-2G183-0000-001 · Tesla P4 (Pascal predecessor) · NVIDIA A2 (Ampere successor) · L4 (Ada)
This accelerator is graded Seller Refurbished. Grading criteria are on How We Test / Grading; warranty and RMA terms on Warranty & RMA.
Refurbishment & test process
- Visual inspection of shroud, PCB, bracket and PCIe fingers.
- vBIOS version and device ID read and recorded.
- GPU compute and full 16 GB GDDR6 exercised under sustained load with ECC error monitoring.
- Thermal soak on a GPU-airflow bench; clocks and power draw verified within spec.
Warranty
365-day owned warranty with seller-paid return freight and a zero restocking fee.
The card ships small-parcel in anti-static, foam-lined packaging with 1–2 business-day handling. Full terms: Shipping & Returns.
- Handling: 1–2 business days · free ground shipping in the contiguous US.
- Volume: multi-card / fleet orders quoted via RFQ.
- Returns: 30-day window, seller-paid return freight, zero restocking fee.