NVIDIA T4
TENSOR CORE GPU
GPU Acceleration Goes Mainstream
NVIDIA T4 enterprise GPUs supercharge the world’s most
trusted mainstream servers, easily fitting into standard data
center infrastructures. Its low-profile, 70-watt (W) design
is powered by NVIDIA Turing™ Tensor Cores, delivering SPECIFICATIONS
GPU Architecture NVIDIA Turing
revolutionary multi-precision performance to accelerate a wide NVIDIA Turing Tensor
320
Cores
range of modern applications, including machine learning, deep
NVIDIA CUDA Cores
®
2,560
learning, and virtual desktops. This advanced GPU is packaged Single-Precision 8.1 TFLOPS
in an energy-efficient 70 W, small PCIe form factor, optimized for Mixed-Precision
65 TFLOPS
(FP16/FP32)
maximum utility in enterprise data centers and the cloud. INT8 130 TOPS
INT4 260 TOPS
GPU Memory 16 GB GDDR6
300 GB/sec
Inference Performance
ECC Yes
GNMT 36X Interconnect
32 GB/sec
ResNet-50 27X Bandwidth
DeepSpeech2 21X System Interface x16 PCIe Gen3
CPU Form Factor Low-Profile PCIe
0 5X 10X 15X 20X 25X 30X 35X 40X Thermal Solution Passive
Comparisons made of one NVIDIA Tesla T4 GPU and servers with a dual-socket Xeon Gold 6140 CPU. Compute APIs CUDA, NVIDIA TensorRT™,
ONNX
Training Performance
ResNet-50 9.3X
CPU
0 1X 2X 3X 4X 5X 6X 7X 8X 9X 10X
Comparison made between dual NVIDIA Tesla T4 GPUs and servers with a dual-socket Xeon Gold 6140 CPU.
NVIDIA T4 | DataSheet | Mar19
Performance to Drive Data Center Acceleration
The small-form-factor, 70-watt (W) design The NVIDIA T4 data center GPU is the
makes NVIDIA T4 optimized for scale-out ideal universal accelerator for distributed
servers, providing an incredible 50X higher computing environments. Revolutionary
energy efficiency compared to CPUs, multi-precision performance accelerates
drastically reducing operational costs. deep learning and machine learning
In the last two years, NVIDIA’s inference training and inference, video transcoding,
platform has increased efficiency by over and virtual desktops. NVIDIA T4 supports
10X, and it remains the most energy- all AI frameworks and network types,
efficient solution for distributed AI training delivering dramatic performance and
and inference. efficiency that maximize the utility of at-
scale deployments.
Turing Tensor Core technology with Turing’s powerful RT Cores, combined
multi-precision computing for AI powers with NVIDIA RTX™ technology, enable
breakthrough performance from FP32 to real-time ray-traced rendering, delivering
FP16 to INT8, as well as INT4 precisions. photorealistic objects and environments
It delivers up to 9.3X higher performance with physically accurate shadows,
than CPUs on training and up to 36X on reflections, and refractions.
inference.
To learn more about the NVIDIA T4, visit www.nvidia.com/T4
© 2019 NVIDIA Corporation. All rights reserved. NVIDIA, the NVIDIA logo, NVIDIA Turing, CUDA, and TensorRT are trademarks and/or
registered trademarks of NVIDIA Corporation in the U.S. and other countries. All other trademarks and copyrights are the property of
their respective owners. Mar19
