'The more GPU you buy, the more you save.'
NVIDIA GTC2018 will always be the beginning of the opening game. Huang Renxun, in the opening speech tonight of the GeForce GTX 2080 in the domestic media, there is no mention of the game and consumer-related products, but it is the above sentence. Always run through two and a half hours of speeches, which is half an hour more than scheduled speeches.
And almost as long as the speech on the commercial level, Huang Renxun is willing to spend more time seems to have become a practice.
The key point is that no matter what you say, new products still have to find ways to sell them. Big brother, 60,000 yuan graphics card, 2.5 million host computer to understand?
Quadro GV100: Redefining tactical graphics
After screaming that performance requirements have soared, NVIDIA RTX ray tracing technology can allow a stand-alone desktop to replace supercomputers to achieve a variety of complex light effects or even Star Wars self-produced short films. The Quadro GV100 has emerged as a professional-grade top-level graphics card. In the hands of Huang Renxun.
This is a possession ECC-protected 32GB HBM2 memory, 5120 CUDA cores, 640 Tensor The professional graphics card, the live demonstration of "Star Wars" Demo is exactly two Quadro GV100 connected through NVLink2, the final host to complete.
This also means that a computer can have 64GB of HBM2 memory and 10,240 CUDA cores, which are used to replace traditional rendering farms in exchange for higher performance and space savings.
But also because it is a professional graphics card, The price also reached 8999 US dollars, nearly 6 million soft sisters.
In contrast, the previously released TITAN V can be described as amazingly cost-effective. However, it cannot be denied that this is currently the highest NVIDIA memory capacity graphics card. The earlier GP100 was limited to 16GB, while the Quadro P6000 was only 24GB.
more importantly This is the only Quadro graphics card that uses the Volta architecture and is the only hardware card that supports hardware acceleration of NVIDIA TRX ray tracing technology.
However, for animation rendering and production studios, it would be even more ideal to use the performance Quadro GV100 instead of a large host.
DGX-2: Give you a reason to spend 2.5 million yuan
About DGX-1's upgraded product DGX-2, Huang Renxun used 'The world's largest GPU As an adjective, this confidence comes from NVIDIA's closer serialization of the Tesla V100 through a technology called NVswitch, which achieves 300GB/s inter-chip transfer efficiency over traditional PCIe 12x transfer speeds, with the help of NVLink2. Combine graphics cards to form a channel with a total bandwidth of up to 14TB/s.
What is the concept of this? In the words of Huang Renxun, If calculated by a 1GB movie, 14,000 movies can be transmitted in one second and a mysterious smile will appear immediately.
of course, This is also likely to be the first GPU that Huang Renxun ever held in one hand.
But DGX-2 is really not a joke. NVswitch is connected to a total of 16 Tesla V100 GPUs, and then paired with a pair of Xeon Platinum, with system memory up to 1.5TB DDR4, storage capacity is 30TB NVMe SSD, of course, the entire system power consumption Reached a terrifying 10,000 watts. It is hardly imagined that a parameter could be placed on a consumer computer.
According to the propaganda, DGX-2's performance is equivalent to twice that of DGX-1, and it can provide about 2 PFLOPs of computing performance. Because of this, pricing is $399,000, and the selling price of 2.5 million softcoins is not an exaggeration.
Continuing to develop GPUs Based on the features of GPUs, Huang Renxun did not hold back a flirtation of Intel at the scene. This time they chose the Skylake CPU as a comparison. Here is the screen where the CPU can process the number of flowers at the same time.
Here is the Tesla V100.
If using Kubernetes technology, the same performance machine can also achieve more simultaneous processing capabilities.
In addition, NVIDIA also responded to Uber’s automatic driving accident last week. They rethought the training process of autopilot technology. They will first test and control the simulation environment, and realize the human’s real events through massive simulations. In various situations, NVIDIA's own SIM artificial intelligence constantly confesses itself, and accident verification that was not thought of in actual experiments can be performed.
At the end of the speech, Huang Renxun also demonstrated a live demo of a vehicle remotely controlled by a VR device.
This time, Huang Renxun put the GTC2018's focus on rendering, artificial intelligence, autopilot, and new platforms. Engineers spend more effort on excavating existing GPU hardware performance than graphics cards and mainframes.
Obviously, in the professional field, NVIDIA's opponents are no longer limited to AMD. They may need to cooperate and compete with depots, medical device manufacturers, and industrial robot manufacturers. The next-generation gaming video cards that we care about may still need to wait patiently. It's time.
