In a nutshell, NVIDIA Pascal architecture is a general purpose built architecture to aid in the computing of computers that implement Machine learning, computer vision and real world simulations via the GPU.
This relatively new architecture is maintained by NVIDIA and is supported natively on the new CUDA Toolkit 8 framework and selection of high performance and computing development tools.
Highlights of the NVIDIA Pascal Architecture
The NVIDIA Pascal architecture leverages the technological advancements right from the GPU hardware to the software that rides on top of it. These optimizations allows this architecture o squeeze out every bit of performance in an efficient manner.
One such system is the world’s first deep learning supercomputer. The NVIDIA DGX-1. This is a fully integrated hardware and software solution with minimal configuration and deployment requirements to get it off to a good start.
The ideal GPU architectures for leveraging the full ecosystem of the Pascal architecture are the NVIDIA Maxwell architecture GPUs and the NVIDIA Kepler architecture GPUs as they provide lots of application performance improvements through shared memory.
The NVIDIA Pascal architecture boast high energy efficiency with its 16 Nanometer FinFET fabrication technology containing 150 billion transistors making the Pascal GPU the world’s largest FinFET chip.
Again the Pascal architecture boasts over 5 TeraFLOPS of double precision High performance Computing workloads with 12 times neural network training performance.
Again the technology uses the NVIDIA NVLink high speed bidirectional interconnect thereby scaling applications across multiple GPUs where applicable.
Processing and data is unified into a single package to deliver high compute efficiency with the Pascal architecture in addition to new artificial intelligence algorithms. This gives it a boost in deep learning and other computational workloads with more than 21 TeraFLOPS of peak performance.
The first Pascal architecture based hardware device is the NVIDIA Tesla P100 accelerator. It comes with an 18 billion transistor Pascal GPU. it is this accelerator that is used in the NVIDIA DGX-1 AI Supercomputer. The supercomputer contains 8 of these Pascal accelerators.