OpenAI, the nonprofit artificial intelligence (AI) research company that is backed by Elon Musk, Peter Thiel, and others, will officially become the first customer of Nvidia’s DGX-1, the “world’s first deep learning supercomputer in a box.” The $129,000 system could help the company accelerate its research.
Generally, OpenAI works on breakthroughs in artificial intelligence, specifically around open source AI that could benefit humanity. Recently, though, OpenAI posted about four special projects it is working on and called for experts to submit applications. Those projects are as follows:
- Detect if someone is using a covert breakthrough AI system in the world.
- Build an agent to win online programming competitions.
- Cyber-security defense.
- Create a complex simulation with many long-lived agents.
With those projects in mind, one can see how OpenAI could make use some additional computing power. That’s where the The Nvidia DGX-1 comes in. Nvidia’s website describes it as “the world’s first purpose-built system for deep learning with fully integrated hardware and software that can be deployed quickly and easily.”
Deep learning is a branch of machine learning, which itself is a form of AI. As a part of deep learning, a system is “trained” on a dataset using an algorithm to determine features of that data and make predictions. For example, Google’s DeepMind was trained on data from matches of the game Go, in order to compete against a master Go player.
SEE: Quick glossary: Artificial intelligence (Tech Pro Research)
One of the core value additions of the DGX-1, Nvidia claims, is that it makes training time up to 75x faster than other solutions. This could mean that AI systems could be tested, and potentially brought to market, faster. The company also claims that it delivers 56x more CPU performance as well, with capability of up to 170 teraflops.
The DGX-1 has eight Tesla P100 GPUs, with 16GB of memory per GPU. It uses the Nvidia Pascal microarchitecture, which is the successor to the Maxwell microarchitecture. In terms of system memory it has 512 GB 2133 MHz DDR4, and storage is made up of four 1.92 TB solid state drives.
The software stack for the DGX-1 includes the NVIDIA Deep Learning SDK, the DIGITS GPU training system, drivers, and CUDA, which is used to help design deep neural networks. It runs Ubuntu Server Linux OS.
In a video posted on YouTube OpenAI research director Ilya Sutskever said that the Nvidia DGX-1 will help improve the firm’s research overall, will shorten some experiments “by weeks,” and will allow them to ask better questions. While many models will be impacted, there are a few that will be improved more quickly.
“Most immediately, our language models that are trained on conversational data and our generative models of images will become much better as a result of DGX-1,” Sutskever said. “We won’t need to write any new code. We will take out existing code and we’ll just increase the size of the model and we’ll get much better results than we have right now.”
The 3 big takeaways for TechRepublic readers
- The Elon Musk-backed nonprofit, OpenAI, will be the first recipient of the Nvidia DGX-1 supercomputer in a box, which could accelerate their AI research.
- The DGX-1 costs $129,000, but could accelerate deep learning training time by 75x and improve CPU performance by 56x.
- Most immediately, the DGX-1 will improve OpenAI’s language and image models, and will cut down the time of some experiments by weeks.