Semiconductor brand new Enterprise GROQ has always been quite low-key, the site does not have any promotional materials, is considered a rather mysterious new enterprise, the outside world is more familiar with it from Google Ai dedicated processor TPU (Tensor Processing Units) team dug most key members, GROQ recently said on the website that it plans to release its first new product in 2018, but it is not yet aware of the shipping time. GROQ said its chip execution speed of floating-point operation (flops) 400 trillion times, is Google's new version of TPU more than twice times, TPU in the depth of learning training stage support 180 trillion times per second (Teraops) operations. GROQ also said its chip performs 8 trillion operations per watt. GROQ was founded by venture capitalists Chamath Palihapitiya less than two years ago, with $10.3 million in investment, and 8 of the 10 members of the team were members of the TPU design team, including Jonathan Ross, one of the founders. Recently Groq also hired Xilinx (Xilinx) business Vice President Krishna Rangasayee as operating director. GROQ in the establishment of less than 2 years to release the first chip, the completion of the development time will be almost equal to Google, Google engineers in a short period of 14 months to create TPU, 1 years after the completion of the 2nd generation TPU. AI Market is now the semiconductor industry strategists battleground, GROQ not only to learn the depth of the hardware mixer Nvidia as a rival, it seems to want to challenge Google and Intel (Intel). In fact, not just GROQ, each chip factory targets Nvidia, and according to Nvidia, most large server makers and cloud operators now use the new Volta architecture to build the GPU. However, the Volta graphics chip uses a hundreds of-tensor core, processing speed of 120 trillion times per second, such choreography is very expensive, CEO Huang revealed that nvidia in the Volta architecture cost about $3 billion. Other rivals are also raising $ hundreds of millions of to build advanced AI chips, trying to shorten the gap with Nvidia. Among them, the Wave computing plans to develop a coarse-grained reconfigurable processor (coarse-grained reconfigurable array) architecture for $60 million; Cerebras Systems raised $112 million and the company's value was estimated to have swelled to $860 million; Graphcore also just got a $50 million financing from VC Sequoia Capital, a company that claims its chips are intelligent processor (IPU) training and inference faster than Nvidia's previous Pascal architecture $number times faster.
