Facebook ‘Big Basin’ AI Compute Platform Adopts NVIDIA Tesla P100 For Next Gen Data Centers
Big Basin was built in collaboration with Quanta Clout Technology (QCT), Facebook's ODM partner. As currently constructed, Big Basin features no less than eight NVIDIA Tesla P100 GPU accelerators. They're connected using NVIDIA NVLink to form an eight-GPU hybrid cube mesh, which is similar to the architecture used by NVIDIA's DGX-1 systems. Combined with the NVIDIA Deep Learning SDK, Big Basin is able to utilize the new architecture and interconnect to improve deep learning across all GPUs.
The new GPU server will allow Facebook to train machine learning models that are 30 percent larger compared to Big Sur. That is a result of having more arithmetic throughput and a memory increase from 12GB to 16GB to play with. The benefit for researchers and engineers is they can move more quickly in developing increasingly complex AI models that will help the social network better understand text, photos, and videos on its platform, and ultimately make better predictions based on that content.
In terms of raw power, the Tesla P100 GPUs give Big Basin a boost in single-precision floating-point arithmetic performance of 10.6 teraflops per GPU, up from 7 teraflops in Big Sur. Half-precision will also be introduced with the new architecture powering Big Basin.
"In tests with the popular image classification model ResNet-50, we were able to reach almost 100 percent improvement in throughput compared with Big Sur, allowing us to experiment faster and work with more complex models than before," Facebook said.
Facebook is not hoarding this technology all for itself. Instead, the company is open-sourcing the design of Big Basin through the Open Compute Project. It plans to release a comprehensive set of hardware design files in the near-future in hopes that it will spur the development of complex AI systems that can be sued to build a more open and connected world.