At Nvidia’s March 2013 GPU Technology Conference, the company announced a breakthrough graphics processing unit (GPU) code-named “Volta” with nearly four times the bandwidth than its current top-of-the-line Kepler graphics cards.
However, at the 2014 GPU Technology Conference, Nvidia changed things around a bit, by placing Volta out more than two years, or well after the 2016 release of its Volta-like “Pascal” GPU’s. Essentially, Pascal will have mostly the same speed and bandwidth characteristics promised for Volta, with a new twist—Nvidia’s own homegrown bus.
If Nvidia’s beastly new Titan X graphics card wows you—and it should—well, hold onto your pants: The company said it’s next-gen GPU technology, code-named Pascal, will offer roughly ten times the performance of the Titan X, at least in compute tasks. The next-generation GPU from Nvidia would also offer a host of new technologies, including 3D-stacked memory and NVLink, said CEO Jen-Hsun Huang during the GPU Technology Conference’s keynote in San Jose. The GPU will essentially be a super computer, Huang said, and Nvidia is shooting for it to be ten times faster than today’s Maxwell-based GPU’s.
Just as the Volta GPU was expected to, Pascal will have “3D stacked memory”—memory chips stacked one on top of the other and joined by circuitry, resulting in graphics cards that can access up to 1 terabyte per second (TB’s) of bandwidth. The 1TB/s threshold will be achieved by stacking memory on top of the GPU itself, with a silica substrate between them, and then cutting a hole through the silicon to connect each layer. Nvidia’s current GPU’s can access only 288 GB/s; upcoming Pascal graphics cards, then, will be about 3.5 times faster. Huang also revealed that a Pascal GPU could run up to 32 GB of RAM, rather than the maximum of 12 GB a Maxwell-based card can use.
This GPU works with Convolution (compute) 4x (FP16) Mixed Precision, Fully Connected (bandwidth) 6x with 3D memory on the forward path. The backward path works with fully connected (bandwidth) 6x with 3D memory and convolution (compute) 4x Mixed Precision.This GPU also has a weight update – interconnect, that is to say, at 10x with NVLINK.With 3x memory bandwidth, we’re pumped to see this next generation in graphics action. A 32GB GPU with 3D memory is just what’s needed to drive the next generation forward – and upward.
Take a peek at our GTC 2015 tag portal as well as our Gaming Hub for more information on all the happenings in this graphics space through the week.