At the end of March, NVIDIA held the GTC 2018 conference (GPU Technology Conference) in American San José. Among others, novelties in the field of computing using GPU accelerators, machine learning, and AI were presented at the conference.
Indeed, this year’s conference offered a lot of novelties. To those interested, we recommend watching the entire GTC 2018 keynote video prepared by the co-founder and general director of NVIDIA, Mr. Jensen Huang – traditionally dressed in a leather biker jacket:
What were we intrigued by?
At the GTC 2018 conference, NVIDIA presented the successor of their flag graphic accelerator for the field of machine learning – NVIDIA V100 with a 32GB memory (which is twice the capacity of the original card). Another novelty is a complex system for machine learning — the NVIDIA DGX-2 server. It is the successor of the first concept by NVIDIA within this field (DGX-1). DGX-2 concentrates the performance of sixteen newest Tesla V100 32GB accelerators into a single server, which was also presented at the GTC conference.
Tesla V100 32GB GP GPU accelerator parameters:
- 5 120 CUDA cores, 640 Tensor cores
- performance 7,8 FP64 TFLOPS, 15,7 FP32 TFLOPS, 125 Tensor TFLOPS (NVLink version)
- 32GB HBM2 @ 900GB/s
- 300GB/s NVLink or PCIe gen3
Parameters of NVIDIA DGX-2, which boasts the most powerful machine learning system in the world!
-
NVIDIA DGX-2 power 2 PFLOPS!
- 16× Tesla V100 32GB GPU
- NVSwitch: 2.4 TB/s bisection bandwidth
- 24× GPU-GPU Bandwidth
- 0.5 TB of Unified GPU Memory
- 10× higher performance for deep learning tasks
Other interesting news and observations from the GTC 2018 conference
Furthermore, NVIDIA presented a support to the Kubernetes technology for the NVIDIA GPU accelerators (for the NVIDIA containers, NVIDIA GPU Cloud — AWS-WC2, GCP, Azure, DGX). The support to this technology significantly simplifies the application of the AI models in the environment of a hybrid cloud.
Another interesting thing was the extension of the support to applications running in the containers above the NVIDIA GPU cloud or an HPC cloud for HPC visualisation. The combination of NVIDIA DGX-1, Pure Storage, and Arista switch – known as AIRI (i.e. AI ready infrastructure) was also very interesting.
GTC Europe 2018, 9.-11. října 2018
And we look forward ti the GTC EUROPE 2018 conference, whitch will be held on 9-11 October 2018 in Munich, Germany. A už se těšíme na konferenci GTC EUROPE 2018, která se bude konat 9.-11. října 2018 v německém Mnichově.
