GPU compared to FPGA advantage to adapt to the rapidly changing needs of AI

　　<strong> Dumbo Reuters </ strong> (the beginning of the source / text) last week, during GTC19 Assembly, NVIDIA accelerated computing Paresh Kharya director of product management for time on the GPU compared to the advantages of FPGA answer questions indicate, the GPU can have the obvious advantage on the program, the entire development time is shorter.
　　He said said the current do a FPGA, the total programming time is coming a few months, but also to program it at the hardware level. Now, however, AI speed changes very fast, even minute update is calculated, it is necessary to implement a highly flexible software-programmable terminal. GPU is precisely the field of AI ASIC, his instruction set is very advantageous, is fully programmable, and is software-defined.
　　Another advantage is that the GPU architecture is forward compatible with the new hardware if future needs, can greatly shorten the development cycle, the entire hardware can be constantly updated with software to adapt software can be updated directly in the library. Meanwhile NVIDIA platform can be used on any device, including desktops, laptops, servers, data centers, and an edge of things.
　　In the keynote speech GTC19 General Assembly, Jen-Hsun Huang also continue to emphasize the obvious advantages in a variety of applications GPU compared to the CPU, such as Ali's limbic system running on the GPU can do 780 queries per second, but if you say with the CPU, can only do 3 times per second query; in Baidu using performance NVIDIA AI recommended system point of view, the past for Baidu huge potential user interest data model training package on costly and slow CPU, GPU and training CPU cost only 1/10, and supports larger model training.