Nvidia releases the world's largest GPU, priced at $400,000. How about mining?

At GTC this morning, NVIDIA released the world's largest GPU: DGX-2. DGX-2 is priced at $399,000 and can achieve up to 20 million floating point operations per second. The following is the real appearance of DGX-2 that we took at the scene. It was also one of the "single products" that attracted the most onlookers that day.

How about using such an awesome GPU to mine ETH?

Please see the detailed report below

On the morning of March 27th, US time, the most important part of NVIDIA's GTC conference - the keynote speech by NVIDIA founder Huang Renxun was held in San Jose.

You know, the GTC conference has always focused on AI and deep learning, and this time is no exception! So what did Huang Xiaoming, who loves leather jackets and even changed into a new one, say?

From Star Wars, oh no, from the ray tracing RTX technology used in it, to the first Quadro GV100 GPU using Volta architecture, and then from the new version of Tesla V100 with 32G memory upgrade, to the supercomputer DGX-2. Yes, DGX-2 is the supercomputer equipped with "the world's largest GPU" in Huang Xiaoming's mouth, which can achieve 200 million floating-point operations per second, consumes 10 kilowatts of power, and is 10 times more powerful than the DGX-1 released by Nvidia last year.

I know that if you are not an expert, you will definitely not know what this means, but you will understand it by giving an example. Using this supercomputer, 14,000 movies can be downloaded per second. How about that? Isn't it amazing?

Of course, NVIDIA has also launched various new versions of machine learning application platforms, next-generation autonomous driving chips, autonomous driving test platforms, etc. I won’t say much about these new terms that have made the industry curious and shocked. Today, let’s talk about what changes these new technologies of NVIDIA will bring if applied to daily life? What application scenarios are suitable? Don’t worry, even a novice can understand it!

Real-time light and shadow tracking technology: opening up a new world for the entertainment industry

Although this Huang Xiaoming is not the same Huang Xiaoming, the new products announced by him this time can benefit many people in the entertainment and animation industry.

NVIDIA's new Quadro GV100 GPU uses RTX technology for real-time ray tracing, which will provide the animation industry with a more efficient way to produce 3D graphics and scenes. Going a step further, practitioners in the media and entertainment industry are happy because they can create realistic animation effects faster; and audiences and players are also blessed because the visual feast will be more and better.

Nvidia showed a Star Wars video at the conference to demonstrate how good RTX real-time light tracing can be: every time a beam of light encounters a surface, the computer must quickly decide how to deal with the relationship between the two: is it reflected or absorbed? From which angle is it reflected? To what extent is it absorbed?

There is no doubt that this kind of light and shadow processing requires a huge amount of calculation. In today's animated films, it takes several hours to render a single scene. But Huang Renxun said that we have entered an era of real-time light and shadow processing, "Everything you see here is real-time." This is why Nvidia dares to call its achievement "the biggest advancement in computer graphics since the introduction of programmable shaders about 20 years ago."

Let's take a look at the parameters section.

Last week, Nvidia revealed that its technology supports more than 24 professional design and creative applications with a total user base of more than 25 million.

The Quadro GV100 GPU, which uses NVIDIA's NVLink interconnect technology, has 32GB of memory, expandable to 64GB with multiple Quadro GPUs, making it the highest performance platform available on the market for these applications.

Based on Nvidia's Volta GPU architecture, GV100 can provide double-precision 7.4 teraflops (teraflops is "floating point operation", which is a measure of the computing power of a computer. A teraflop is one trillion floating point operations per second, so GV100's computing capacity is extremely powerful), single-precision 14.8 teraflops and deep learning performance 118.5 teraflops. The Nvidia OptiX AI-denoiser built into Nvidia RTX can achieve nearly 100 times the CPU performance, enabling real-time noise-free rendering.

Huang Renxun himself was also very excited

And then, how did the entertainment industry respond?

That would definitely be welcome.

The CEO of Epic Games, which produces the "Fortnite" and "Infinity Blade" series, said in a statement: "The advent of Nvidia RTX technology has brought real-time light and shadow tracing into the next chapter. By providing powerful technology to the game development community with support for the new DirectX Raytracing API, Nvidia has become the driving force behind the next generation of game and movie graphics." The words are more official, and the key point is "For game development, this is truly a groundbreaking technology."

Remedy Entertainment said: "After developing with NVIDIA RTX technology, we were amazed at how fast it was and how much higher visual fidelity it had than traditional technology. We are excited to think about what we can achieve with RTX in the future - it's time to prepare something special for players!"

After talking about entertainment, let’s take a look at medical care.

At the conference, Huang Renxun demonstrated Clara, Nvidia's first supercomputer dedicated to medical image processing. What's so great about this supercomputer?

Remember those medical images taken by ultrasound? Most of them are 2D and black and white. But as long as the 2D black and white image data is transferred to the Clara computer and processed by artificial intelligence software, medical images can provide more information. Colors, layers, and regions can be rendered in real time on the original black and white images. In other words, expectant mothers can see a 3D heart and the appearance of the baby in the womb.

What's more practical is that hospitals can directly connect this computer to existing medical equipment without replacing it. At present, NVIDIA is working with many medical manufacturers, including GE General Electric, Samsung Electronics, etc., as well as AI medical startups such as Tuma Shenwei and Infervision. In addition, Infervision is the first artificial intelligence company in China to apply deep learning technology to medical imaging diagnosis.

It is conceivable how huge the market would be if NVIDIA's super medical imaging computer were deployed in major tertiary hospitals.

The supercomputer equipped with the world's largest GPU

The most important thing about NVIDIA's conference today is the release of what is known as the "world's largest GPU". Let's take a look at the predecessor of this supercomputer DGX-2, DGX-1.

At the 2016 NVIDIA GPU Global Technology Conference, NVIDIA launched the world's first deep learning supercomputer DGX-1, which is the first system designed specifically for deep learning and provides a throughput equivalent to 250 x86 servers. How powerful is the computing power? It is equivalent to putting 250 servers in this box.

So what improvements does DGX-2 have over DGX-1?

First of all, it has 16 Volta GPUs, which gives DGX-2 the deep learning processing power of 300 servers. From other parameters, there is a total of 512GB HBM2 memory, which can provide up to 14.4TB/s throughput and 81920 CUDA cores.

But these 16 GPUs are not simply connected, because DGX-2 is the first system to debut NVSwitch, which enables 16 GPUs in the system to share a unified memory space, which enables professional developers to process the largest data sets and the most complex deep learning models. This NVIDIA Volta GPU connected by NVSwitch can be said to have created the world's largest GPU.

Secondly, it is faster. DGX-2 can train FAIRSeq, a state-of-the-art neural machine translation model, in less than two days. This is a 10-fold performance improvement over DGX-1. Speaking of speed, we also mentioned it before, because it is up to 14.4TB per second, 14,000 movies can be downloaded per second.

The world's most powerful GPU is priced at $399,000 (about RMB 2.5 million) and will be available in the third quarter of this year. At this price, I think I can't afford it, but how much money can it save for enterprises? The price of 300 dual-CPU servers is about $3 million, and Nvidia's price is only 1/8 of the cost. No wonder Huang Xiaoming's catchphrase at the GTC conference today was "the more you buy, the more you save."

Autonomous driving: Testing suspended, research and development continues

At the end of the press conference, Huang Renxun said that Nvidia would suspend road testing of autonomous vehicles, but research and development would continue.

At the CES conference on January 8 this year, Nvidia officially announced its cooperation with Uber, and its chips will become an important driving force for Uber's fleet. In fact, Uber has been using Nvidia's technology since it deployed the Volvo test fleet in 2016. Before the Uber accident last week, Nvidia tested its driverless cars in New Jersey, Santa Clara, California, Japan and Germany. After the accident, Nvidia announced that it would suspend its driverless car testing worldwide, and its stock price immediately fell.

There is no precedent for the peaceful breakup of automakers and hardware suppliers: In May 2016, Tesla's autopilot system "Autopilot" made a mistake due to multiple reasons, causing a man driving a Tesla Model S to run under a container truck without slowing down, and the driver died on the spot. Shortly thereafter, Tesla's camera supplier at the time, Israeli company Mobileye, announced the termination of cooperation with Tesla. Although the two parties did not give much explanation, the outside world generally believed that the accident was an important reason for the termination of cooperation.

Of course, the driverless car test is only suspended. Earlier, a spokesperson for Nvidia responded to this matter by saying: "In the future, the safety of driverless cars will be far greater than that of human drivers, so the research and development of driverless cars needs to continue. But in order to learn lessons from the Uber accident, we will suspend the test."

“As long as it can move, it will become autonomous driving”

However, today's press conference shows that Nvidia is not going to stop exploring autonomous driving: Nvidia plans to build a new system Drive Sim and Constellation to test autonomous vehicles, including:

.AV Verification System

.VRAV Simulator

.Same architecture as DRIVE computer

.Simulate rare or difficult conditions, recreate scenarios, run regression tests, and accumulate virtual testing mileage

The first part of the DriveSimandConstellation system is called DriveSim. DriveSim is a software platform that can simulate the sensors used in unmanned vehicles. It runs on a modular hardware platform, with each module consisting of 8 high-end graphics processors. Since modules can be added at any time as needed, this method can simulate all the sensors on the vehicle. The information processed by the simulated sensors can be so realistic that it is almost impossible to distinguish with the naked eye.

NVIDIA DriveConstellation

Using data recorded on the road, NVIDIA GPUs can change the position of the sun, weather, road reflectivity, etc. When simulating scenes, summer noon, night driving, downpours or blizzards can all be switched at will.

All of this is designed to run in data centers, which isn’t cheap, but of course, it’s safer to test in simulation than to risk your life in the real world, and millions of miles can be accumulated every day under infinitely varying conditions.

The Drive series is part of NVIDIA's Pegasus artificial intelligence computing platform. Samples of the Pegasus motherboard will be available later this year, but the simulator can also run on any NVIDIA Drive automation platform. NVIDIA is very proud of the software compatibility of all its hardware: software can be developed on existing hardware first, and then easily ported to new platforms if it needs to be moved to new hardware.

Huang Xiaoming has said so much today, but I didn't expect that Nvidia's stock fell by 7.7%. Is this a good time for beginners to buy Nvidia stocks? What do you think?