Nvidia has claimed that its next-generation GPU architecture, codenamed Pascal, will offer a ten-fold boost to computer science tasks like deep learning.
In a presentation at the company's GPU Technology Conference this week, Nvidia co-founder and chief executive officer Jen-Hsun Huang discussed the upcoming architecture while positioning its GeForce GTX Titan X consumer card as being suitable for computer science tasks as well as Ultra HD gaming. Shortly after unveiling a neural-network training platform dubbed DIGITS and a hardware DevBox implementation which includes four Titan X GPUs, Huang revealed that Pascal will include vastly improved performance over the current Maxwell architecture for such projects.
'[Pascal] will benefit from a billion dollars worth of refinement because of R&D done over the last three years,
' Huang told attendees at the event, before detailing three design features that will be of interest to those working in neural network fields: support for mixed-precision computing, helping alleviate the performance hit that comes from switching from single- to double-precision float compute; the NVLink high-speed interconnect, claimed to be five to 12 times faster than PCI Express, for multi-GPU implementations; and card models with up to 32GB of VRAM.
Huang also confirmed that Pascal will include stacked memory, placed adjacent to the GPU, to improve power efficiency and accelerate performance over the current planar Maxwell design. In total, Huang claimed, the shift to 3D memory in Pascal will mean a three-fold increase in memory bandwidth and a near-three-fold increase in maximum VRAM per GPU - hence the jump from the 12GB framebuffer seen in Titan X to Pascal's upper limit of 32GB.
While Huang's presentation focused on Pascal's suitability for neural network simulation and other machine-learning applications, the same architecture will power its future consumer cards. As a result, expect to see the next-generation GeForce family boasting all of these features - although perhaps not 32GB consumer-oriented card models, at least initially.
More information on NVLink is available from Nvidia's official product page
, while its DIGITS deep-learning platform can be downloaded here