NVIDIA founder and CEO Huang Renxun released HGX-2 at the recently held GTC in Taipei, a "building block" cloud server platform that enables server manufacturers to create more powerful systems around NVIDIA GPUs for high-performance computing. And artificial intelligence.
The platform's performance has increased 500-fold in five years and is supported by ecosystems including all computer manufacturers and ISVs.
Huang Renxun said that computing systems will bring changes to the trillion-dollar industry worldwide, and GPUs are at the heart of the computing ecosystem. He believes that a series of breakthroughs in materials science, energy and medicine depend on higher computing power.
“The computing needs are unprecedented, we need to expand Moore's Law,” Huang said.
Huang Renxun detailed the GPU-based deep learning-driven technology "Cambrian Big Bang." In less than a decade, GPU computing power has grown 20 times, or 1.7 times a year, far exceeding Moore's Law.
However, due to the development of AI, our demand for performance is still "increasing (but not slowing down)." “Before, the software was written by humans, and software engineers could only write limited software, but the machine would not be exhausted.”
"As long as there is data, as long as we know how to create the architecture, we can build powerful software," Huang Renxun said. “Every company that develops software around the world needs an AI supercomputer.”
NVIDIA HGX-2 combines breakthrough features such as the NVIDIA NVSwitch interconnect fabric to seamlessly connect 16 NVIDIA Tesla V100 Tensor Core GPUs into a single "mega GPU." NVIDIA partners will launch their first HGX-2-based systems later this year.
“With the cooperation with you, anyone can use this future fusion computing method to create high-performance computing and AI applications,” Huang Renxun expressed his gratitude to the partners in the entire computer industry. "We have a variety of servers."
Industry partners use NVIDIA's server platform extensively
The core of HGX-2 isNVIDIA Tesla V100 GPUIt is equipped with 32GB of high-bandwidth memory capacity and provides deep learning performance of 125 trillion floating point calculations per second.
BYNVSwitchConnecting up to 16 Tesla V100 GPUs through NVSwitch creates the “world's largest GPU” in Huang Renxun.
"Each GPU can communicate with other GPUs at 300GB/s bandwidth and 10 times PCI Express at the same time," Huang said. "That means all GPUs can talk to each other at the same time."
Huang Renxun also introduced the new model in detail.NVIDIA DGX-2This is the first system built using the HGX-2 server platform. It weighs 350 pounds and provides 2 petaflops per second and 512GB of HBM2 memory.
"This is the fastest single computer ever created: an operating system, a programming model that you can program as a computer," Huang said. “It’s like a personal computer, but it’s quite fast.”
New rules of the industry
Huang Renxun said that the DGX-2 achieved a leap of ten times the computing performance of the previous generation in just six months. In the AlexNet image recognition benchmark, a 500-fold leap was achieved compared to the performance achieved with a pair of NVIDIA GPUs five years ago.
“The industry has already ushered in new rules,” said Huang Renxun. "This new algorithm tells us: 'If you can do it, and if you are willing to optimize it across the stack, you can achieve a very fast performance boost."
The result: a series of deep learning speed records emerged from the top of the technology stack, from single-chip performance to large data center systems.
NVIDIA delivers record-breaking performance across the technology stack
NVIDIA GPU Cloud (NGC) Extends the accessibility of this performance. It allows researchers to no longer be limited to desktop and server systems, but to leverage cloud systems from providers such as Amazon, Google, Alibaba and Oracle.
“Every layer of this software has been tuned and tested,” Huang Renxun added, and now 20,000 companies have already downloaded NGC.
The results of the real world are amazing. Huang Renxun, along with an NVIDIA AI researcher, demonstrated the details of removing photos in real time, such as trees or street lights.
Keep in mind the principles of the PLASTER
Huang Renxun said that in addition to the demonstration, to deploy such a deep learning application on a large scale, it is necessary to solve seven challenges: programmability, latency, accuracy, size, throughput. Throughput, energy efficiency, and rate of learning. Use the English initials of these seven words as an abbreviationPLASTER。
The ability to scale will be the key to bringing a new generation of artificial intelligence services to action, a process called reasoning for speech synthesis and recognition, image and video processing, and referral services.
To demonstrate this expansion capability, Huang Renxun demonstrated an amazing floral recognition system that can achieve a speed of 4 images per second based on a CPU and a maximum of 2,500 images per second with a single GPU. Reaching this speed four times, thanks to passingKubernetesAdd more GPUs to support more desktops, data centers and cloud service providers.
"We call it Kubernetes (Nubdianets on NVIDIA GPUs) on the NVIDIA GPU, referred to as KONG," said Huang Renxun.
With Kubernetes, GPU computing can reach the maximum size you need
"Kubernetes can be described as a very large-scale operating system," Huang Renxun said. "If you look at the entire stack, from the GPU to all the APIs and libraries that can be put into Docker and containers and run on Kubernetes, the software stack is so complex that hundreds of engineers have been working around for years. ”
NVIDIA servers will provide superior computing power to the multi-trillion dollar industry.
In the $2 trillion entertainment industry, NVIDIA is newRTXTechnology accelerates ray tracing, providing cinema-quality quality standards, plus real-time graphics technology and artificial intelligence that will accelerate today's traditional CPU-based rendering farms.
In the $7 trillion healthcare industry, ourProject ClaraIt is expected that GPU computing will redefine medical imaging, based on current medical instruments to achieve higher fidelity, or use fewer rays to generate images of the same quality (reducing the scanner's radiation by one-sixth, Safe for children).
Safe cities also represent another $2 trillion market opportunity.
NVIDIA is working through its end-to-endDRIVE platformSolve problems in the transportation sector. Transportation represents another $10 trillion market. “Everything that can move in the future will be automated,” Huang Renxun detailed NVIDIA's ability to collect and process data, train models, simulate billions of miles of driving, and drive real.
"zoom out" and teleport into the micro space
At the end of the speech, Huang Renxun used virtual reality technology to “shrink” one of our colleagues and teleported it into a mini-car, and then the colleague drove the car in a miniature city.。
That is to say, with such technology, human beings can become backup forces of AI machines in virtual space, no matter where they are or how large they are.
“In the future, you will be able to blend in with the robot,” Huang Renxun told more than 2,000 developers, researchers, government officials and the media. “You will be able to achieve telepresence and reach whatever you want. local."