[Netease smart news May 11 news] Nvidia held this year's GPU Developer Conference, the company CEO Huang Renxun keynote speech lasted more than two hours, its content involves all aspects of machine learning, and based on the Nvidia graphics processing chip (GPU ) Many applications. The highlight of the speech was its announcement that Nvidia had released the new Volta architecture and the corresponding architecture V100 chip.
Nvidia has been working hard to make its graphics processing chip more suitable for artificial intelligence applications, and added 16-bit floating-point fast calculations to the chip. But its new Volta architecture takes this specialized function to a new level, and by adopting the new tensor core, it greatly accelerates the training and reasoning ability of the neural network. Volta's tensor core is designed specifically for neural networks, while the traditional GPU core is mainly used to perform classic graphics operations such as shadow processing.
For neural networks, the basic building blocks are matrix multiplication and addition. Nvidia's new tensor core can perform all operations and can perform multiple sets of dual 4x4 matrix operations while adding a third matrix parallel computation. Therefore, 5120 cores can be run in parallel on the V100, and each core itself can run other operations in parallel. Nvidia said that the result is that in the inferential learning process, V100 is 12 times faster than Pascal, and the reasoning speed is 6 times that of Pascal.
The NVIDIA V100 is one of the most impressive chips ever made. In an area of ​​815 square millimeters, it is covered with 21 billion transistors. The CEO of Nvidia said that it is the largest and most complex chip that can be manufactured in the semiconductor field. The chip's R&D cost is 3 billion U.S. dollars, while the chip is manufactured by TSMC's 12nm process technology and uses Samsung's high-speed memory.
After the keynote speech, Nvidia explained that the reason why it uses the large size of 12nm is to create the most advanced chip. Google also recently created a customized TensorFlow chip for artificial intelligence, but the release of Volta architecture chips will effectively prevent its competition.
Volta is clearly part of the NVIDIA artificial intelligence strategy, but it has not stopped there. Huang also announced TensorRT at the conference, which is the Tensorflow and Caffe compilers, designed to optimize gpu performance. The compiler not only improves efficiency, but also greatly reduces the deficiencies of Google's custom chips. This makes the TensorFlow chip 30% slower than the Skylake or P100 image recognition benchmark. In terms of equipment load, the Tesla V100 PCIe can replace more than a dozen traditional CPUs and consume much less energy.
In addition, Nvidia also made a more direct response to custom reasoning chip competition, and announced that the company is developing DLA (deep learning accelerator and code open source software. Its tensor core configures up to 20MB of register files, up to 16GB of HBM2 memory Read and write speeds of up to 900GB/s, and NVLink input and output up to 300GB/s, the result is that the Volta architecture is exactly the chip architecture built for artificial intelligence.
Nvidia later confirmed that not all Volta architecture processors have such high artificial intelligence acceleration capabilities, and some may be more focused on pure graphics or general-purpose computing performance. Nvidia explains that its tensor core is ideal for training and reasoning operations, not to create a single product line.
The V100 chip will be the core processing chip of the supercomputers DGX-1 and HGX-1. It is reported that the upgraded DGX-1 uses eight V100 chips, will be released in the third quarter, priced at about $ 149,000. The DGX Station with 4 V100 chips sells for $69,000 and is also planned for release in the third quarter. The V100-based OEM products are expected to start shipping before the end of this year.
Nvidia also collaborated with Microsoft Azure to develop a cloud-friendly supercomputer HGX-1 with eight V100 chips that can be flexibly configured to meet various cloud computing needs. Microsoft plans to use the Volta architecture on its own applications and open it to Azure users.
Nvidia expects that in addition to pure software applications, Volta will also provide support for self-driving cars and robots. Nvidia expects that Volta architecture-based processors and circuit boards will become the core of artificial intelligence devices for learning and reasoning.
This includes a variety of robot systems, especially those using Isaac's newly released Isaac robot simulation tools, as well as various types of self-driving cars. Among them Airbus is designing a small aircraft that can take off vertically and can carry two passengers to 70 miles. It also uses the new Isaac robot simulation tool released by Nvidia.
(English source /entremetech compiler / machine Xiaoyi proofread / æ™— ice)
Car Phone Wireless Charging Coil,Circuit Board Induction Coils,Intersection Induction Coils,Induction Coil Set
Shenzhen Sichuangge Magneto-electric Co. , Ltd , https://www.rodinductor.com