Nvidia announces the Volta architecture

This year’s edition of the GPU Technology Conference (GTC 2013) begins entirely in the tradition of Nvidia’s CEO Jen-Hsun Huang, who discusses new advances in the GPGPU area – the ability to use graphics processors for general computing. In addition, the company’s latest roadmap is presented with the upcoming architectures Maxwell and Volta.

The sequel to today’s Kepler is called Maxwell. The details of the new architecture are few, but Huang reveals that unified virtual memory is one of the main ingredients. One of the most time consuming tasks for a modern graphics processor is to move data between the CPU and GPU, something that can be largely eliminated when both processors see what the other is storing in memory. The launch of Maxwell will take place at the beginning of 2014, which indicates that the circuits are manufactured with TSMC’s 20 nanometer technology.

Next in line is Maxwell’s successor and tonight’s big news: Volta. After Maxwell solved the shared memory problem, Volta will deliver higher bandwidth through stacked memory, where at least part of the memory is placed above the GPU. The lack of memory bandwidth is said to be the biggest bottleneck in parallel calculations with a graphics processor.

The stacked memory is based on the Through-Silicon Via (TSV) technology, which means that several memory circuits can be joined in a common capsule and communicate with each other through vertical conductor paths. The result is lower energy consumption and significantly higher bandwidth because the total line distance to the graphics processor is reduced. Together, the enhancements will provide access to massive one terabyte per second of memory bandwidth – compared to 288 GB / s for the Geforce GTX Titan and Radeon HD 7970 Gigahertz Edition.

However, Nvidia’s planning does not state when Volta is intended to find its way into the market, although the diagram may indicate 2016.