Nvidia is going to introduce their much-anticipated fifth-generation Pascal GPU architecture primarily targeted at the HPC and AI markets later today at the GTC 2016.
While we were waiting to hear the key details about the new hardware from Nvidia co-founder and CEO Jen-Hsun Huang himself later in the day, some seminars at the event had already divulged a lot of info on the overall performance enhancement Pascal is expected to bring when it finally hits the market.
Apparently, a slide featuring about the Nvidia cuDNN improvement was spotted by the Videocardz team (via Hardware.fr). The slide, which was meant for a session on artificial intelligence at the ongoing GTC 2016, mentioned Nvidia’s plans to boost the general performance through the updated CUDA Deep Neural Network (cuDNN) library.
The new cuDNN v5 library update includes:
1# Accelerated deep learning: Caffe, Tenorflow, CNTK, Torch, and Therano.
2# High-performance Deep Neural Network training, and
3# Gradually increasing performance over time.
Being the state-of-the-art next-generation graphics architecture from the chip giant, Pascal flaunts a number of advanced technologies including, HBM2, NVLINK, and Mixed Precision. The leaked slide also displays that Pascal, complemented by the cuDNN v5 can offer a performance factor up to 12 times higher than what we have previously seen.
Worth noting, the Maxwell-based Tesla M40 with as well as the cuDNN v3 and Kepler-powered Tesla K40 with cuDNN can deliver a six and four times higher performance enhancement respectively. With all that into consideration, the updated library and Pascal collectively promises a 70% performance enhancement as compared to Maxwell.