Nvidia got a lot of gamers excited when it unveiled the GeForce GTX 970 last September, and for good reason: It's a powerhouse of a graphics card for just $330. We were duly impressed with the 970 after pitting it against some AMD GPUs and Nvidia's own higher-priced GTX 980 in a graphics card slugfest last fall.
But gamers have since complained that the GTX 970 seems to falter under pressure when a game tries to use more than 3.5GB of the card's 4GB of VRAM.
Nvidia recently addressed these complaints and admitted the performance issues weren't the result of overzealous imaginations, a firmware flaw, or a few bad units rolling off the production line. The problem was the way Nvidia designed the memory for the GTX 970. It also turns out that some of Nvidia's originally reported specs for the GTX 970 were wrong.
Why this matters: The revelations about the GTX 970 leave Nvidia in an embarrassing situation. Not only does it have to account for why some edge case-users may be experiencing performance problems, but the company had to come clean about giving reviewers incorrect specs, and not correcting them for months. It's a problem for Nvidia, but it's more important that anyone looking at graphics cards understands the limitations of the GTX 970—limitations that should've been made public from the start.
Although the 970 has 4GB of VRAM, not all of that memory is created equal. The 970 has two different segments of memory it can access, as Nvidia recently explained to PC Perspective: a primary segment with 3.5GB and a secondary, slower one with another 512MB.
Under normal conditions (and the vast majority of gaming scenarios), the card relies on the 3.5GB segment alone, turning to the supplementary segment only when necessary. But when a game needs more than 3.5GB of memory, tapping into the secondary segment drags down performance. Some are reporting stutters and frame rate drops as the card uses between 3.5GB and 4GB of memory.
Another issue, as reported by AnandTech, is that the GTX 970 came with fewer render output units (ROPs) than originally claimed. ROPs are directly tied to a card's performance and the GTX 970 was reported as having 64, when it actually has 56. The GTX 970 also has a smaller L2 cache than originally stated, at 1.75MB instead of 2MB. Nvidia told AnandTech the discrepancy was the result of an error by the technical marketing team rather than an attempt at duplicity.
What it means to you
Despite these shortcomings, however, the GTX 970 still delivers an amazing amount of bang for your buck. The difference now, as AnandTech points out, is that we have a better understanding of the card's strengths and weaknesses. While the issues introduced by the memory design may only be bumped into in edge cases, they may sway gamers who want to play at very high resolutions and/or with anti-aliasing settings cranked to the max to consider AMD's Radeon cards, as those situations require more memory than typical gaming scenarios.
The flagship Radeon R9 290X delivers performance similar to the GTX 970 in most cases, includes 4GB to 8GB of memory with a wider 512-bit memory bus (as opposed to the GTX 970's 256-bit bus), and can often be found for less than the $330 and up Nvidia card—and sometimes far less if you can find juicy rebates. That makes the Radeon better suited for situations that require more memory—again, very high resolutions and/or cranked anti-aliasing settings—but AMD's cards are far less energy efficient than Nvidia's. Decisions, decisions.
And surprise! There's already a Change.org petition by upset GTX 970 buyers demanding a refund from Nvidia.
For more in-depth details about the GeForce GTX 970's memory details, be sure to check out AnandTech's in-depth examination of the issue, which spans multiple pages and is chock full of nitty-gritty details.