Nvidia DGX-2 is the "world's largest GPU"
Unfortunately no next-gen GeForce GPUs were unveiled at GTC 2018 yesterday, but Nvidia did present the DGX-2 GPU server that is essentially the world’s most powerful AI system. Thanks to the improved NVSwitch scalable architecture, the DGX-2 can also pass as “the world’s largest GPU”, as it interconnects 16 Tesla V100 GPUs comprising 81,920 CUDA cores.
There are two stacked boards, each supporting 8 Tesla V100 GPUs. The setup cumulates 512 GB of HBM2 video memory and the GPUs are interconnected via 12 high speed NVSwitches that offer 2.4 TB/s bisection bandwidth. The systems is able to deliver 14.4 TB/s aggregate bandwidth, while the interconnect bandwidth link each GPU can process up to 300 GB/s.
Since the DGX-2 is actually a server, it also comes with eight EDR Infiniband 100 GbE adapters that are compatible with 10/25 GbEthernet connections. Additionally, the server part is powered by two Intel Xeon Platinum CPUs coupled with 1.5 TB system memory and NVMe SSDs that amount to 30 TB of storage.
Compared to the previous DGX-1 model introduced in September 2017, the DGX-2 can train the FAIRSeq neural machine translation model 10 times faster (in less than two days). The new GPU rack consumes 10 KW, weighs around 350 lbs and can be ordered for US$399,000.