At the GPU Technology Conference in Japan, Nvidia Corp. once again revealed key features of its next-generation graphics processing architecture code-named “Pascal”. As a it appears, the company has slightly changed its plans concerning memory capacity supported by its upcoming GPUs.
As expected, Nvidia’s high-end graphics processor that belongs to the “Pascal” family will feature an all-new architecture with a number of exclusive innovations, including mixed precision (for the first time Nvidia’s stream processors will support FP16, FP32 and FP64 precision), NVLink interconnection technology for supercomputers and multi-GPU configurations, unified memory addressing as well as support for second-generation high-bandwidth memory (HBM generation 2).
Based on a slide that Nvidia demonstrated at the GTC Japan 2015, next-generation high-end graphics cards with “Pascal” GPUs will sport up to 16GB of HBM2 with up to 1TB/s bandwidth. Previously Nvidia expected select solutions with “Pascal” graphics processors to feature up to 32GB of HBM2.
Given the fact that Nvidia does not produce high-bandwidth memory itself, but relies on supplies from companies like Samsung Electronics and SK Hynix, changes of their roadmaps can affect Nvidia’s plans. In order to install 32GB of HBM2 memory on a graphics processor with a 4096-bit memory bus, 8GB memory chips are used. While the HBM2 specification allows to build such ICs [integrated circuits], it is not easy to manufacture packages with eight vertically stacked 8Gb memory dies. As a result, such chips may be delayed from 2016 to a later date.
16GB of HBM2 memory should be enough for gaming and professional graphics cards, but high-performance computing applications could take advantage of 32GB of onboard memory even now.
Discuss on our Facebook page, HERE.
KitGuru Says: If Nvidia is not able to get 8GB HBM2 chips next year, AMD will not be able to get them as well. Therefore, expect graphics cards with up to 16GB of high-bandwidth memory from both companies.