Nvidia has bombarded the market in the GPU Technology Conference 2017. As per reports, Nvidia has announced bigger than biggest GPUs, strictly speaking, the best ever processing units in the world. One has to decide if it is a GPU or any beast processing core. Nvidia has announced the features, applications and many other technical specifications of its Volta's Tensor V100 chip. In simple words, Volta’s Tensor Cores are to neural networks what traditional GPU cores are to graphics.
Matrices are the major operation inputs to GPUs. The Nvidia's V100 chip can perform a 4x4 matrix multiplication and along with the addition of another matrix simultaneously. The V100 is capable of running 5120 core at a time. Each core, in turn, is capable of executing millions of operations. The result of this high-performance chip, is speeding up in inferencing learning over Pascal.
Coming to the fabrication specifications, the V100 chip consists of 21 billion transistors, with 12nm technology. The chip was erected on a die of 815 square millimetres. In terms of cost, the fabrication of this chip costed 3 billion USD starting from R&D.
The Tensor cores are complemented with a large 20MB register file, 16GB of HBM2 RAM at 900GB/s, and 300GB/s NVLink for IO. The major application of this chip is targeted at Artificial Intelligence and Machine Learning.
Nvidia has announced DGX-1 built with eight V100 chips, which is priced at $149,000 USD. Another variant with four V100 chips quoted at $69,000 USD. Nvidia has developed HGX-1, which is targeted in cloud computing. As Microsoft is venturing into the complex technologies like Mixed Reality and automobile infotainment, Microsoft is making partnership with Nvidia. Also, Microsoft is working on applications specific for Artificial Intelligence and Machine Learning supporting the DGX-1 and the HGX-1 processors. Nvidia is planing to release both these processors in Q3 of 2017 and products based on these processors from OEMs are expected to release by the end of the year.
Matrices are the major operation inputs to GPUs. The Nvidia's V100 chip can perform a 4x4 matrix multiplication and along with the addition of another matrix simultaneously. The V100 is capable of running 5120 core at a time. Each core, in turn, is capable of executing millions of operations. The result of this high-performance chip, is speeding up in inferencing learning over Pascal.
The Tensor cores are complemented with a large 20MB register file, 16GB of HBM2 RAM at 900GB/s, and 300GB/s NVLink for IO. The major application of this chip is targeted at Artificial Intelligence and Machine Learning.
Nvidia has announced DGX-1 built with eight V100 chips, which is priced at $149,000 USD. Another variant with four V100 chips quoted at $69,000 USD. Nvidia has developed HGX-1, which is targeted in cloud computing. As Microsoft is venturing into the complex technologies like Mixed Reality and automobile infotainment, Microsoft is making partnership with Nvidia. Also, Microsoft is working on applications specific for Artificial Intelligence and Machine Learning supporting the DGX-1 and the HGX-1 processors. Nvidia is planing to release both these processors in Q3 of 2017 and products based on these processors from OEMs are expected to release by the end of the year.
