NVIDIA's technical conference "GTC 2018" opens, introduces a small supercomputer "DGX-2" composed of over 80,000 CUDA cores and virtual simulation technology for automatic driving cars

Technology conference "GPU Technology Conference 2018" held by semiconductor solution maker · NVIDIAGTC 2018) Will be held at the San Jose McEnery Convention Center in San Jose, California, USA from 26 to 29 March, 2018 local time. In this conference where the technology for automatic driving and AI was emphasized, CEO Jensen Juan of NVIDIA entered the presentation on the second day and made a presentation, with 80,1920CUDA CoreSuch as DGX-2, a small supercomputer that says 'the world's largest GPU', and a virtual simulator 'DRIVE Constellation' that lets you run automated driving cars for billions of kilometers in a virtual world. We are announcing what we should do.

NVIDIA Reinvents the Workstation with Real-Time Ray Tracing | NVIDIA Newsroom
https://nvidianews.nvidia.com/news/nvidia-reinvents-the-workstation-with-real-time-ray-tracing

◆ Compact supercomputer "DGX-2" with a price of over 40 million yen, CUDA Core of 80 1920 units
With GTC 2018 NVIDIA is using the data center GPU "Tesla V 100"Which is 32 GB which is doubled from the conventional memory capacity"Tesla V100 32GB GPUWe announced. And it is the supercomputer's new product "DGX - 2" that interconnects 16 GPU units and operates as one "world's largest GPU" by connecting with a bandwidth of 2.4 TB per second.

DGX-2
https://www.nvidia.com/en-us/data-center/dgx-2/?ncid=pa-tra-d2lh-34943


The DGX - 2 secures a total of 512 GB of memory capacity by connecting 12 Tesla V100 32 GB with 12 NVSwitch, a newly developed NVLink - based switch. And Mr. Fan said that "realizing half-precision floating point calculation performance of 2 PFLOPS is a major feature with" only 10 kW power consumption ".



The price is announced at 399,000 dollars (about 42 million yen). It is not a priceless price, but Mr. Fan emphasizes the high cost performance as well. If you try to realize equivalent performance on a CPU basis, its cost will be 3 million dollars (approx. 320 million yen), the power consumption will reach 180 kW, but with DGX-2 you can get 300 racks of 15 racks With a deep learning processing capability comparable to that of a server, the size is one sixteenth that of the server, and the power efficiency has 18 times the small space · high efficiency.


When Mr. Hwang announced DXG-2 at GTC 2018, "The tremendous enhancement of this deep learning is only a small part of the functions that will be announced in the future. We are based on NVIDIA's deep learning platform which became the world standard to the world.The company dramatically improves the performance of the platform greatly exceeding the Moore's Law and it is dramatically improving the performance of medical, transportation, scientific exploration and other countless fields We will bring breakthroughs that will encourage change in. "

◆ "DRIVE Constellation" simulation system that allows safe driving of billions of kilometers of automated driving car in a virtual environment
"DRIVE Constellation" cloud-based virtual simulation system developed for testing automated driving cars using photorealistic simulation reproduces the environment closely resembling the real world in a computer, runs a virtual automatic driving car in it Making it possible to advance the learning of on-board AI without using real vehicles. This means that a safer and more extensible method for introducing automatic driving vehicles used for development on public roads will be created.

This system is a computing platform based on two separate servers. The first server runs "NVIDIA DRIVE Sim software" and simulates various sensors installed in automatic driving cars such as cameras, Lidar and radar. And the second server equipped with a powerful AI in-vehicle computer "NVIDIA DRIVE Pegasus" executed from a sensor of a car running on an actual road by executing a set of software stacks for automatic driving vehicles Process as if it is data.


Rob Cheong-gaa, vice president and general manager of NVIDIA's automotive business, said about this technology, "To introduce production-level automatic driving vehicles, it is necessary to run billions of miles (billions of km) We need a solution to acquire the safety and reliability that our customers need, and by combining NVIDIA's expertise in visual computing with the knowledge of data centers, DRIVE Constellation By using virtual simulation, we conducted tests on billions of miles (billions of kilometers) of custom scenarios and troublesome rare cases and compared the robustness of the algorithm The time and expense to spend all these are just the same on the actual road Considering the testing, it is only a minor amount, "he says in the press release.

NVIDIA GPU is installed in the simulation server, each GPU generates simulated sensor data and sends it to DRIVE Pegasus for processing. Driving instructions from DRIVE Pegasus are fed back to the simulator, and a digital feedback loop is completed. This "hardware-in-the-loop" cycle is done 30 times per second and is used to verify that "algorithms and software running on Pegasus are correctly manipulating simulated vehicles" I will.

DRIVE Sim software generates a photo realistic data stream and creates enormous numbers of different test environments. This software can simulate, for example, abnormal weather such as storm and snowstorm, dazzling sunlight in various time zones during the day, limited visibility at night, all kinds of road surface and topography. Hazardous situations are described in the simulation and you can test the ability of an automotive vehicle to respond without risking humans.


DRIVE Constellation is expected to be offered for early access partners in the third quarter of 2018.

◆ Graphic board "Quadro GV100" that enables "real time ray tracing"
Graphic board that makes it possible to realize more realistic CG generation "Ray tracing" by real time calculation by reproducing the appearance of light in the eye with the same mechanism as the natural environment "Quadro GV 100 "has also been announced. Using NVIDIA RTX technology,Volta architectureQuadro GV100 based on 5120 CUDA cores, 640 Tensor Core, 32 GB HMB 2 memory, 7.4 TFLOPS in double precision (FP 64), 14.8 TFLOPS in single precision (FP 32), 118.5 TFLOPS in Tensor performance The ability to generate realistic graphics based on physical characteristics is provided by having the computational power of.


The Quadro GV100 can be easily implemented using various APIs, "Realistic lighting, reflection, and shadowing using actual light and physical properties" "Significant improvement in rendering performance by AI" "NVLink Extendable expandability of memory up to 64 GB using "Can be collaborated, designed, and created with immersive VR" and so on. It is already possible to purchase at the official store etc., and the price at the time of article creation is 8,999 dollars (about 950,000 yen).

Buy Professional Graphics Cards & Workstations | NVIDIA Quadro
https://www.nvidia.com/en-us/design-visualization/quadro-store/


"NVIDIA employs ray tracing technology optimized for the Volta architecture and combines it with high-performance hardware to create a work piece," said Bob Petty, vice president of professional visualization business at NVIDIA. The transformation of the station has been realized.The experts and designers can simulate and manipulate the creation in a way that was impossible so far.This makes it possible for the workflow of various industries to be fundamentally It will change from that. "

◆ Alliance with semiconductor manufacturer ARM, integration of the latest development environment "TensorRT4" to Google's TensorFlow 1.7
At the conference, NVIDIA also partnered with SoftBank's semiconductor manufacturer "ARM" to provide deep learning inference for billions of mobile devices, home appliances, and Internet (IoT) devices to be introduced into the world market It is announced that it will go.

In this tie-up, NVIDIA and ARM will incorporate the open-source NVIDIA Deep Learning Accelerator Architecture into ARM's "Project Trillium" platform to realize machine learning. This collaboration is aimed at allowing IoT chip companies to easily incorporate AI into their designs and to deliver intelligent, low cost products to billions of consumers around the world .

In addition, a series of new technologies and partnerships have been announced that extend the capabilities of deep learning reasoning to hyper-scale data centers and at the same time greatly reduce the cost of providing services using deep learning. At the same time, a new version of TensorRT inference software and the integration of TensorRT into Google's "TensorFlow framework" widely used in the world are announced. Mr. Ian Buck, NVIDIA's vice president and general manager of accelerated computing said, "Using GPU acceleration for production deep learning reasoning, real-time and lowest cost for large-scale neural networks We can increase the quality of deep learning and quicken the support for more intelligent applications and frameworks to improve the quality of deep learning and reduce the cost of 30 million hyper-scale servers It is now possible to contribute. "