NVIDIA introduced the supercomputer DGX-2, which the company calls “the world’s largest GPU.” The machine is able to work with 16 Tesla V100 modules with 32 GB of HBM2 memory, one of which was shown to visitors of GTC 2018. The company also talked about NVSwitch and NVLink2, with which all graphics accelerators are combined into a single unit, providing a throughput of 14.4 TB / from.
The supercomputer supports up to 1.5 terabytes of RAM, up to 60 terabytes of NVMe-drives. The power consumption of the system running on the Intel Xeon Platinum pair (28-core chips are mentioned) will be 10 kW.
The performance of NVIDIA DGX-2 is about 2 petaflops. It can be used to train neural networks and other similar tasks. The price of the issue is $ 400 thousand, the pre-order can be made on the company’s website.