NVIDIA Announces New GPU – Volta: A Redesigned GPU Architecture Customized For AI
May 11th, 2017: NVIDIA had its annual GPU technology Conference or GTC recently. The company took a giant step towards an accelerator processor architecture that has been specifically designed for Artificial Intelligence or AI. NVIDIA recently announced the Volta: a GPU architecture customized for deep learning and AI.
This new development of artificial networks was made possible by deep learning through NVIDIA's GPUs and the CUDA software platform. NVIDIA holds a major share of the deep learning networks that are in use today. The GPU giant is in collaboration with over 40,000 companies and 0.5 million developers for advancement in neural network applications.
NVIDIA announced the next generation of GPU architecture at the GTC this year. The new architecture has been named "Volta". While the chip might still be referred to as a GPU, Volta is much more capable than that. The company has added 640 new tensor cores that are capable of processing in 4x4x4 matrix multiples. This mathcore helps for additional processing in deep learning environments.
Developers will be able to take advantage of tensor cores for their AI framework without having to rewrite their applications, all thanks to NVIDIA SDK libraries and runtimes like cuDNN and TensorRT.
Volta has been announced with over 5000 GPU CUDA cores and over 300 GB of system communications Bandwidth through six high-speed NV links and 16 GBs of second generation high bandwidth memory, also known as HBM2 .
This means that the new Volta architecture will sit in the same power envelope and form factor as the previous Pascal architecture but with an extra 1.5x the memory performance and 2x the NVLink performance and a total of 120 teraflops of processing performance with the Tessar cores. For more information on NVIDIA Volta, visit here.
