Manufacturer's description and specifications:

NVIDIA Tesla P4 GPU - System Overview

Performance

Using a low-profile design, the P4 is perfect for smaller blade servers. As a direct successor to its Maxwell counterpart, the NVIDIA Tesla M4 GPU, the Tesla P4 packs 7.2 billion transistors on a large chip with a die area of 314 mm. A server with a single Tesla P4 replaces 13 CPU-only video-inferencing servers, which means theres a decrease in total cost of ownership. The NVIDIA Tesla P4 supports 2560 CUDA cores, 160 texture mapping units, and 64 ROPs.

Memory

Under the hood, the NVIDIA P4 has a memory clock running at 1502MHz, or 6 Gbps. Also known as the GP104 GPU, this graphics processing unit uses 8GB of GDDR5 memory connected to a 256-bit memory interface with a 192.3GB/s memory bandwidth. The GPU also operates at a frequency of 886MHz, which can be boosted to a maximum of 1114MHz.

Features

The NVIDIA Tesla P4 GPU has redefined visual computing by offering developers, scientists, and artists the ability to take on the largest visualization challenges with immersive, interactive, and photorealistic worlds. Outfitted with NVIDIA Quadro vDWS, professional designers can work remotely and access applications and data from anywhere, on any device. To release the heat on this 75W TDP card, there is a passive heatsink with bidirectional airflow. Power consumption is minimal, and the card does not require an additional power connector, helping make it 40x more energy efficient than other GPUs.

Summary

The NVIDIA Tesla P4 GPU accelerator delivers new software for massive leaps in efficiency and enhanced speed to accelerate inferencing production workloads. These GPUs are ideal for artificial intelligence services such as voice-activated applications. Offering 8GB of GDDR5 memory and a low 75W power supply, the P4 is an innovative, powerful, and cost-efficient graphics card that can help users access data from anywhere, on any device.

Based on a 16nm manufacturing process, this card is primarily used for professional graphics, deep learning inference, and virtualization. The NVIDIA Tesla P4 GPU is also packed with NVIDIA Quadro Virtual Data Center Workstation (Quadro vDWS) software to allow full GPU-per-server density.

NVIDIA Tesla P4 GPU - Specifications

• NVIDIA CUDA Cores: 2,560
• TMUs: 160
• ROPs: 64
• SM Count: 20
• L1 Cache: 48KB (per SM)
• L2 Cache: 2MB

• GPU Memory: 8GB GDDR5
• Memory Interface: 256-bit
• Memory Bandwidth: Up to 192.3 GB/s

• FP16 (half) performance: 89.12 GFLOPS
• FP32 (float) performance: 5.704 TFLOPS
• FP16 (double) performance: 178.2 GFLOPS
• Transistor Count: 7,200 million

• Base Clock: 886 MHz
• Boost Clock: 1114 MHz
• Memory Clock: 1502 MHz/6 Gbps effective

• Architecture: Pascal
• GP104 graphics processor
• Fabrication Process: 16nm

• PCI Express 3.0 x 16

• 75W
• Power connector: None

• 2.1 H 6.6 L
• Low profile

• None

• DirectX 12.1
• CUDA 6.1
• Vulkan 1.2
• OpenGL 4.6
• OpenCL 1.2
• Shader Model 6.4