Nvidia wants graphics processors to create avatars with realistic human faces, and power visual searches in which images can be identified to produce matching search results.
Those are already possible with Nvidia's fastest graphics processors based on the Kepler architecture, which is being used in Titan, the world's fastest supercomputer. Nvidia on Tuesday provided details to Kepler's successors, which feature technology enhancements that will boost the quality of graphics and make it easier for programmers to write GPU-based applications.
Nvidia CEO Jen-Hsun Huang announced the new graphics processors, code-named Maxwell and Volta, as part of a product roadmap update during the company's GPU Technology Conference in San Jose, California.
Nvidia offers a range of graphics processors including GeForce for PCs, Quadro for workstations and Tesla for servers and supercomputers. Nvidia's GeForce graphics cores are also in its Tegra processors, which are used in tablets and smartphones. The Maxwell and Volta architectures will be used in the upcoming generations of Nvidia's graphics processors.
The biggest enhancement in the Maxwell architecture will be the addition of hooks that could unify GPU and CPU memory. GPU memory will be able to read CPU memory and vice versa, and the shared memory resources will make it easier for developers to write applications, Huang said.
Memory used in graphics processors and PCs are based on different technologies, but can be linked by virtualization technology such as VGX, which is already offered by Nvidia Linking the memory resources directly will make it easier for the GPUs and CPUs to share multiple threads and ensure that workloads and its branches are handled and executed correctly.
Maxwell will also reach mobile devices with Nvidia's upcoming Tegra chip code-named Parker, which will be Nvidia's first 64-bit chip based on the ARM architecture. The Parker chip will have 3D transistors, which should improve performance on mobile devices while preserving battery life.
Following Maxwell will be Volta, which will be an "even more energy-efficient" graphics processor, Huang said.
Volta's size will be smaller due to the implementation of a new chip design, Huang said. Nvidia is stacking DRAM directly on the silicon substrate around the CPU, which is different from the traditional design in which memory was laid flat.
The DRAM will be connected to each other directly, which will result in much faster bandwidth than today's graphics processors, Huang said.
"Volta is going to solve one of the biggest challenges in GPUs today, which is memory bandwidth," Huang said. "We're going to achieve 1 terabyte per second of bandwidth."
The bandwidth is equivalent to taking a Blu-ray movie and moving it through the entire chip in one-fiftieth of a second, Huang said.
The Volta design draws from the idea of 3D stacking, in which chips are stacked on top of each other to reduce the size of a processor.
Nvidia in the past was criticized for its Fermi GPU architecture, which was considered too power hungry. However, Nvidia solved some of Fermi's problems with Kepler, and hopes for the future chips to be faster and more power efficient.
Nvidia also showed a computer called Kayla. which Huang called the world's fastest ARM computer.
Based on Tegra 3 processors, Nvidia said it created a new low-power GPU and combined it with ARM processors to create the computer.
Kayla brings the computing power and GeForce and Tesla to one computer, Huang said. Nvidia said the computer is capable of doing real-time ray tracing, which generates accurate images by tracing paths of light. In addition, the computer also supports CUDA 5, OpenGL and also PhysX.
This was not the first time that Nvidia's Tegra 3 chip is being used for supercomputing. The Barcelona Supercomputing Center in 2011 built a supercomputer using quad-core Tegra 3 processors.