So today was the long awaited day, because during the GPU Tech Conference 2016 it was expected that NVIDIA will reveal new information about its new graphics architecture Pascal.
NVIDIA Pascal is 70% faster than Maxwell in Deep-Learning, also features New NVLink Technology
NVIDIA has shown a new slide at GTC mentioning Pascal, comparing Pascal with Tesla K40 and M40 on a new library for deep-learning. NVIDIA Pascal will rely on a new library CuDNN for deep-learning (CUDA Deep Neural Network) proposed major changes to its new GPU architectures. With cuDNN v1 of Tesla K40 (GK110) it has achieved double performance over previous graphics solutions, while the deep-learning speed of Maxwell M40 has increased 6.25 times. The “deep learning” is one of the main priorities of NVIDIA in gaming performance and Pascal naturally pushes the bar even higher in this area. Therefore, Pascal running cuDNN v5 sees 10.5x boost over the first iteration of cuDNNup. On in other words around 70% faster if we compare it with the flagship GPU core of Maxwell, the GM200.
However, it is difficult to judge the overall performance of this Pascal GPU showcasing unique figures, so we have to wait for NVIDIA to reveal information on the increased raw power of the GPU or specific optimizations of deep learning architecture.
On the other hand, Pascal architecture based GPUs will use the new technology NVLink, a new connection for graphics cards that allow to share information between 5 and 12 times faster than today’s PCI Express ports. As a result, it is expected to eliminate the bottleneck latter generate and enable a new generation of highly scalar supercomputers with processing power up 100 times higher than current.
According to tests conducted by NVIDIA with the first endowed servers technology NVLink, called QuantaPlex T21W-3U and Rackgo X Big Sur with two graphics cards performance improves by 400 percent over the current PCI-Express interface, while in 4-GPU configuration it exceeds 500 percent improvement.