How to find a solution to this impossible problem, Nvidia CEO Jensen Huang (Huang Renxun) gave his answer at the GTC Taiwan 2018 conference held on May 30th.
NVIDIA's platform Yewang
Recalling the history of NVIDIA's development from simply launching GPUs to assisting domestic and foreign manufacturers in building GPUs and then enhancing consumer entertainment experiences for computer games. Although it once lost in the smart phone market, AI has been actively deployed in recent years. Began to the data center, the development of the automotive field.
From the perspective of business architecture, NVIDIA first invested 90% of its R&D expenditures in GPU architecture and CUDA software platform, and then applied this model to various platforms: games, data centers, artificial intelligence, and autonomous driving.
NVIDIA has always been a leader in the GPU market, and the gaming industry is the largest source of GPU users. The data center is its fastest growing business market.
In the past three years, the average annual growth rate of this part of business was 85%, and the growth rate in fiscal year 2018 was 133%. These growth came from various vertical fields such as HPC (high performance computer group), cloud computing companies and AI researchers.
In AI technology, NVIDIA has outperformed other competitors. In June 2017, Nvidia released the next-generation Volta GPU with Tensor cores. Relying on deep learning, Volta is five times faster than its predecessor, Pascal. Volta has been adopted by all mainstream clouds. Calculated by vendors and server manufacturers.
This time through the release of NVIDIA HGX-2, NVIDIA officially merged artificial intelligence and high-performance computing into a single platform.
It is understood that HGX-2, as a cloud server platform, has multi-precision computing capabilities, supports high-accuracy FP64 and FP32 calculations for scientific computing and simulation, and also supports FP16 and INT8 precision for AI training and reasoning. HGX-2 performs AI training The speed achieved 15500 images per second on the ResNet-50 benchmark, equivalent to 300 servers with only CPU.
The GPU computing era has arrived
In this GTC Taiwan, Huang Renxun also stated that the annual demand for computing will increase by 100 times in the next 10 years. At the same time, under the condition that Moore's law is gradually declining, the GPU computing volume of the world’s top 50 supercomputers will be in the future. It will grow 15 times in five years. At the same time, GPU-accelerated computing will become the main mode of extending Moore's Law.
Currently, supercomputers have become important tools for the development of modern science. They play an important role in the development of molecular construction, quantum chemistry, quantum mechanics, weather forecasting, meteorological research, energy exploration, physical simulation, data analysis, and artificial intelligence. One hundred billion or billion operational performance.
According to OpenAI statistics, the artificial intelligence computing model will grow by 300,000 times over the next five years, which is expected to grow 30,000 times faster than Moore's Law. With GPU acceleration, the complexity of data and calculation programs can be greatly increased. This solves the computational needs that people in the past could not solve.
Huang Renxun once again stressed that in the past, NVIDIA created the accelerated benefits of the CUDA computing model. It also shows that the model for GPU-accelerated computing will continue to expand in the future. It is expected that the global computing demand in 2028 will be equivalent to that of 10 million Volta architecture GPUs. Traditionally, supercomputer-level computing capability is formed by stacking multiple CPUs, which will occupy large-scale space and high power consumption. If GPUs are replaced, it will save more space and power consumption, and at the same time bring higher acceleration. .
In addition, Huang Renxun stated that NVIDIA officially launched the world's top ODM partnership program. Global design and manufacturing companies like Hon Hai Precision, Inventec, Quanta Computer and Wistron will all become partners and accelerate the various needs of AI cloud computing.
Unified Computing Platform NVIDIA HGX-2
On May 30, 2018, Nvidia announced the launch of the first unified computing platform NVIDIA HGX-2 for both artificial intelligence and high-performance computing.
From the product line perspective, HGX-2 is an upgraded version of HGX-1 last year. The latter is also a reference architecture for Amazon AWS, Facebook, and Microsoft cloud services. HGX-2 is also outside the cloud service platform and can also be used. Go to HPC (high performance computer), becoming the industry's first standard platform for cross-domain computing applications.
The HGX-2 cloud server platform provides multi-precision computing capabilities that provide unique flexibility and provide powerful support for future computing. Nvidia says the platform can perform high-precision FP64 and FP32 operations for scientific computing and simulation, and target AI training. And reasoning for FP16 and Int8 operations, versatility to meet the needs of today's increasingly integrated applications of HPC and AI.
The core of the HGX-2 uses 16 Volta tensor core GPUs to form a large core group through the NVSwitch interconnect structure. As a single giant GPU, the HGX-2 provides 2 petaflops of AI performance. The first built using HGX-2. The system is the recently released NVIDIA DGX-2.
Huang Renxun pointed out that HGX-2 is part of the NVIDIA GPU acceleration server platform series. This series of products is connected to the entire data center server system and is suitable for each large market. It can recommend the most suitable GPU for different AI, high-performance computing and speeding up operations. Combined with CPU configuration. Such as HGX-T for hyperscale training and HPC; HGX-1 for large-scale inference and intelligent image analysis (IVA); and SCX-E for data center, HPC, IVA, virtual desktop infrastructure ( VDI) etc.
Huang Renxun said: 'The HGX-2 with Tensor Core GPUs has brought the industry a powerful, general-purpose computing platform that can combine HPC and AI to solve the enormous challenges facing the world.'
It is understood that Foxconn, Inventec, Quanta, and Wistron, the world's four largest ODMs, are also designing systems based on HGX-2, which are expected to be put into use in the second half of the year in several of the world's largest cloud data centers.
More than a decade ago, NVIDIA decided to transition from a graphics chip company to a computing company and began to build the infrastructure and ecosystem needed to promote GPU computing.
More than ten years later, the achievements of NVIDIA's transformation have been obvious to all. From the most advanced physics, medical research, to the current hottest artificial intelligence and autopilot research and development, NVIDIA's platform continues to provide developers with higher computing performance. Become an engine to promote scientific and technological progress!