When Nvidia held its yearly GPU technology conference, it was just a gaming company looking for new markets, particularly for its specialized chips. Back then, high-performance computing was the key target.
A change came in 2012 when AlexNet emerged and solved the ImageNet problem, causing a widespread growth of deep neural networks that are trained on GPUs.
Currently, Nvidia’s data center business produces $2 billion in yearly sales leaving bigger rivals trying to catch up while venture capitalists invest their money in AI hardware startups looking forward to creating a better mousetrap.
Nvidia no longer requires making any case for GPU computing. However, chief executive officer Jensen Huang’s job during the 2018 GTC’s opening keynote was to persuade the present 8,500 developers that the company can uphold its edge over its competition and deliver the advantages of AI to a wider audience.
According to Huang, believing that the same processors used in playing games would turn into more general-purpose hardware was one of the best decisions by Nvidia. In fact, the company’s CUDA GPU computing platform has been downloaded over 8 million times. Also, the 50 fastest supercomputers in the world currently depend on GPU accelerators to deliver 370 petaflops of horsepower.
The adoption of GPU computing is growing at a fast rate. The reason behind this scenario is that new interconnects, architectures, systems, and algorithms have allowed GPUs to scale quickly. In the last five years, Moore’s Law has grown the performance of CPU by five times while GPUs have expedited molecular dynamics simulations by a factor 25.
As such, a traditional HPC group with 600 dual-CPU servers that consumes 600 kilowatts can currently be substituted with 30 quad-Tesla V100 servers drawing 40 kilowatts. Huang said that the world requires bigger computers since there is a serious job to be done.
A similar advancement appears to be happening in deep learning. In fact, the AlexNet CNN or convolutional neural network was not only eight layers but also had millions of parameters.
In five years, the world experienced a Cambrian explosion of convolutional neural networks (CNNs), generative adversarial networks (GANs), reinforcement learning, recurrent neural networks (RNNs) and new species like capsule nets. All these neural network models have grown in complexity 500 times to billions of parameters and hundreds of layers, building demand for faster hardware. `
Nvidia solution is its DGX-2, which is a new server that has 16 Tesla V100 equivalents that come with 32G each of stacked HBM2 or rather High-Bandwidth Memory all connected over a high-speed switch. The server also features two Intel Xeon Platinum CPUs, 30TB of NVMe solid-state storage, 1.5TB of system memory as well as 100Gbps Ethernet Networking and InfiniBand EDR.
For Nvidia, Automotive ranks among its most high-profile. The company has made a huge, multi-year bet on end-to-end systems for autonomous driving and ADAS (advanced driver assistance).
Nevertheless, the keynote news revealed that the company had put its autonomous driving program on hold in the advent of Uber’s fatal crash. This situation has pushed Nvidia’s shares down significantly.