NVIDIA announced during the GDC 2018 the new Quadro GV100 professional GPU, the first PCI Express card to feature 32GB of HBM2 memory and to use the GV100 for the first time in solutions such as DGX-1. This new GPU will cost $9,000, tripling the cost of the previously released Titan V.
NVIDIA Introduces New Quadro GV 100, First PCIe Graphics Card with 32GB HBM2
The GV100 is composed of six GPC (Graphics Processing Clusters) with 84 Volta Streaming Multiprocessor Units and 42 TPCs that include two SM each. Each SM has 64 CUDA cores totaling 5120 cores that can be used in FP32 and INT32 calculations, while 2560 cores are also available to perform FP64 (double precision) calculations. Also, the card has 640 Tensor Cores and 320 Texture Units.
On paper, the new Quadro GV100 achieves 7.4 TeraFlops of power in double precision, 14.8 TeraFlops in Single Precision and 29.6 TeraFlops in medium precision. On the AI side, 118.5 DLOPs are executed, a slight increase over the Tesla V100 due to the improved operating frequencies.
The memory has been configured with eight 512-bit controllers that make up a 4096-bit bus for the 32GB of HBM2 stacked memory, to which has been added 768KB of L2 memory that adds up to a total of 6MB of L2 for the chip total. With a memory frequency of 850Mhz, a bandwidth of 870GB/s has been achieved, 150GB/s more than the Quadro Pascal GP100.
The card will arrive on the market with the usual reference model with a blower and a PCB that incorporates four Display Port 1.4 outputs. Regarding power consumption, the TDP adds 250W with two 6-pin and 8-pin PCIe connectors respectively.