Nvidia’s New GPU Architecture is set to debut in 2016
After AMD driving great competition with Nvidia, Nvidia is busy and is working hard for it’s next new GPU architecture “Pascal”. Nvidia’s new GPU architecture “Pascal” is set to debut in 2016. According to Nvidia, Pascal will be 10x faster than it’s current GPU architecture Maxwell, sounds awesome right ? There’s more, Pascal will offer up to whooping 32GB of memory 2.7x more than what Nvidia’s flagship TITAN X Offers. Pascal Will Also Offer NVLink which allows faster data transfer between GPU and CPU. Pascal will also feature mixed-precision computing along with 3D Memory which will result in up to 5x interconnect which is used to link two or more GPUs together which eventually will lead to a total of 10X improvement in deep learning applications.
What mixed-precision computing does ?
Mixed-Precision computing is for greater floating point accuracy. Mixed-Precision computing allows Pascal architecture-based GPUs to compute at 16-bit floating point accuracy at twice the rate of 32-bit floating point accuracy. The increased floating point performance in Pascal benefits in classification and convolution (two key activities in deep learning) while achieving required accuracy.
What 3d memory does ?
Memory bandwith constraints and limits the speed at which data can be delivered to the GPU. What 3D memory will do is that it will provide 3X bandwith and nearly 3X the frame buffer capacity of Maxwell, due to which a huge improvement in performance will be seen, this will also let developers build even larger neural networks and accelerate bandwidth intensive portions of deep learning training.
The memory chips of Pascal will be stacked on top of each other and will be placed adjacent to the GPU instead of further down the processor boards. This will help in reducing the distance that bits needs to travel as they transverse from memory to GPU and back, in result it will dramatically accelerate communication and improve power effenciency.
What Is NVlink ?
NVLink allows pascal to move data between GPUs and CPUs 5-12X faster than they can with today’s current, PCI-Express. NVLink will greatly benefit applications such as deep learning, that needs high inter-GPU communication.
NVLink in Pascal allows double the number of GPUs that we can use today in a system to work together in deep learning computations. Due to NVLink CPUs and GPUs can connect in a while new way enabling more flexibility and energy efficiency in server design as compared to PCI-Express.
Alternate link : http://playit.pk/watch?v=Tb7ZYSTYHbw