Nvidia Corp. on Tuesday introduced its future-generation Volta graphics architecture that is due in 2016 – 2017 timeframe. Among other advantages and innovations that Volta will have, it will support a technology capable of solving the main GPU problem ever: insufficient memory bandwidth. With Volta, Nvidia plans to install ultra high-speed stacked memory chips right onto the substrate with graphics processor.
While today’s graphics processing units are extremely powerful, their performance is limited by memory bandwidth in a number of cases. For example, today’s highest-performing single-GPU graphics card – Nvidia GeForce GTX Titan – has peak memory bandwidth of 288GB/s. Given the progress expected from GDDR5 memory technology, it is unlikely that it will double its performance by 2014, when Nvidia’s next-generation Maxwell architecture arrives, or quadruple in three to four years when Nvidia Volta comes out. However, Nvidia seems to have a solution.
In addition to all the innovations, Volta will support stacked DRAMs (interconnected using through-silicon vias) installed onto the substrate of the graphics processor with around 1TB/s of bandwidth. Since the stacked DRAMs will be located very close to the GPU, it will be possible to access memory at extreme speed, thus providing nearly four times memory bandwidth of the GeForce GTX Titan.
Unfortunately, Nvidia did not reveal too much information about the new Volta graphics processing architecture. It is unknown whether it will only rely on stacked DRAMs or will feature multi-tier memory architecture, which will include both stacked DRAMs for performance and traditional GDDR5 memory chips for capacity.