Posted by Chris Scott Barr on Mar 22, 2013

GTC 2013 – Nvidia Maxwell And Volta Bring Unified Virtual Memory And Stacked DRAM

At this week’s GTC, Nvidia talked about a lot of things that aren’t of interest to most gamers. While their GRID VCA looks fascinating from an enterprise standpoint, it doesn’t really have much to do with gaming. However, there were a few tidbits of information that are worth discussing.

First, CEO Jen-Hsun Huang unveiled the company’s desktop GPU roadmap. As you know, last year Nvidia released their first Kepler-based cards, which were the 600-series. The next architecture is codenamed Maxwell. The biggest feature for Maxwell is use of Unified Virtual Memory. Essentially, this allows both GPU and CPU to access the same virtual memory pool. This should result in higher speeds in a variety of applications, as you’re cutting out the middle-man when your GPU needs to access CPU memory, or vice versa. This sounds similar to what AMD is trying to do with HSA.

While we’ve known about Maxwell for a while, Nvidia gave us a peek at what comes after. Code-named Volta, this new architecture will focus on Stacked DRAM. This means that you can expect memory directly on the GPU die, thus increasing bandwidth significantly. How much more bandwidth? Roughly 1 TB/s. For comparison, the Titan only tops out at 288 GB/s.Or as Huang put it, you would be able to transfer an amount of data equivalent to a Blu-ray disc through the chip in 1/50 of a second.

According to the roadmap, we’ll see Maxwell sometime in 2014. This points to a Kepler refresh this year, which would follow what we’ve seen in the past from Nvidia. If we’re looking at historical data, we can also guess that we’ll see Volta-based cards sometime in 2016.

Post a Comment