In the framework of the GPU Technology Conference (GTC), Nvidia has released several interesting announcements highlighting the new version of its DGX system, which is a kind of personal supercomputer focused on artificial intelligence tasks. In fact, the first model, the DGX-1, was presented to Elon Musk in 2016 for the work of OpenAI.
The Nvidia DGX-2 is approximately 10 times faster than the first version, and now it is capable of offering up to 2 petaFLOPS of power, which is mainly due to the renewed Tesla V100 GPU, with Volta architecture, which now integrates 32GB of memory.
A personal beast for artificial intelligence
This supercomputer based on GPU presents a new NVIDIA NVSwitch, which will allow connections up to 300 GB / s, that is, up to 12 times the speed of PCIe. Inside we have 16 Tesla V100 GPUs grouped in a single system, to have a bandwidth that exceeds 14 TB / s.
Continue Reading: What graphics card to buy for a PC Gaming 2018?
In addition to this, we have a couple of Xeon Platinum CPUs with 1.5 TB of RAM and 30 TB of NVMe storage, expandable to 60GB, in addition to InfiniBand and 100GigE connectivity. This monster has a consumption of 110 kW and a weight of almost 160 kilograms. That’s how it looks when compared to the previous version.
The key to this system is its processing capacity and speed, since the 16 GPUs are grouped in a unified memory space, which allows to maintain workloads constantly without much suffering. For example, Nvidia ensures that the DGX-2 can complete a neural network model for language translation up to 10 times faster than a DGX-1 system , which results in a reduction to less than two days, while the DGX-1 needed 15 days in total.
The Nvidia DGX-2 is aimed at the professional segment of the deep learning market, so its price has been set at $399,000, with a launch date scheduled for the third quarter of this year. Meanwhile, for those wishing to build a powerful system that does not require large workloads, the DGX-1 will be available starting at $ 150,000.