On Tuesday, Nvidia announced its next GPU, called Pascal, which will dramatically rearchitect the traditional graphics card by replacing the PCI Express bus and combining it with a new 3D memory technology.
And for those who need a supercomputer now: Nvidia also announced its next-generation Titan card, the GeForce GTX Titan Z, starting at a whopping $2,999. Nvidia also announced what it calls the Iray VCA: a $50,000 virtual computing appliance for rendering images using modeled photons, to produce photorealistic images at up to 60 times the speed of today's Nvidia-powered workstations. Finally, Nvidia announced the $192 Jetson TK1, an embedded development board to take its Tegra K1 to robots and computer vision, and disclosed the name of its next-generation Tegra chip, dubbed Erista.
Nvidia will ship the Pascal processor in 2016, Nvidia chief executive Jen-Hsun Huang said during the keynote address of its GPU Technology Conference, which focuses on the application of GPUs in computational problems. The conference originally was predicated on Nvidia's CUDA programming language, which allowed scientists to program a GPU much like a computer.
It will replace Maxwell, the graphics architecture Nvidia announced last month.
It's unclear how much Nvidia's work, as evidenced by the GPU Technology Conference, contributes to Nvidia's bottom line. But Nvidia has consistently seen growth for in its high-end, high-margin GPUs for the desktop and workstation, which has been partially offset by declines in PCs and notebooks. And problems that Nvidia can solve with its supercomputer GPUs can trickle down into the PC.
And Nvidia faces a thorny one. The problem, as Huang explained it, is relatively simple to express: GPUs are among the biggest chips ever created, but the demands for increasing GPU performance are unceasing. Accessing the memory used as a scratchpad for GPU computations can be improved by using a wider bus interface, but that adds more pins — and there can be a physical limit to how many pins can surround a chip. Simply forcing more data through those pins at higher and higher clock speeds increases the power consumed — and the heat produced — to unmanageable levels.
The answer, Huang said, was a twofold solution: NVLink, a replacement for PCI Express, which will offer between five and 12 times the performance of PCI Express; and the 3D interface, which will stack memory and other chips on top of one another, running via wires through the chips and the substrate to save space. The latter solution will help push memory bandwidth to 1000 times where they are now, Huang said.
The module that Nvidia built to house the Pascal chip architecture is one-third the size of a typical PCI Express module, Huang said. A new connector will allow it to be plugged into the motherboard, Nvidia said.
Sign up for CIO Asia eNewsletters.