Users' curiosity is inexhaustible, especially when it involves a piece of hardware like GeForce GTX 970 graphics card of Nvidia. Some enthusiasts were unpleasantly surprised to discover that the card reports only 3.5 GB of RAM out of the four gigabytes its specifications indicate. With user backlash and a bit of negative press, Nvidia has begun damage control, sharing multiple technical details about its product.
A Surprising Specification Gap
Hardware specifications should always be handled with great care. A wrong number can give erroneous impressions about a graphics card, a hard drive, or a processor. The user buys the piece with those specifications in mind, and once installed discovers the truth the hard way. Something like that has happened with the GeForce GTX 970. Its official launch was last September, and it presented a series of specifications that place it almost on par with the GTX 980. The only differences were in the number of CUDA cores, texture units, core frequency, and TDP. What happened? The specifications published by Nvidia at that time were wrong.
The Technical Explanation
Long story short: Instead of having 64 rendering output units (or ROPs), the GeForce GTX 970 has 56, the Maxwell streaming modules are 13 and not 16, and its L2 cache is actually 1.75 megabytes. If it were only a typo, it wouldn't be so serious, but the story doesn't end there. Several users have discovered that the graphics card reports 3.5 GB of VRAM allocation instead of the full memory, something that doesn't happen with the GTX 980. This is due to memory segmentation, which separates the first 3.5 GB into a high-priority field, and the remaining 512 megabytes into a low-priority one. At the time of publishing the specifications, the marketing team ignored the ability to disable independent ROPs. The error managed to reach the final revision of the text, passed through specialized press undetected, and escaped all identification tools. It took four months to find it.
This is not the first time Nvidia has applied memory segmentation in its products to manage a limited amount of resources and the reuse of defective chips, and it has never become so mediatic before, which brings us to the fundamental dilemma: Accessing those final 512 megabytes causes a significant performance loss (28 gigabytes per second versus 192 for the main segment). It all comes down to the way software uses the graphics card's memory, but there are many angry users, and others decided to submit a petition on Change.org for Nvidia to enable a refund mechanism for those who bought the card. This is the kind of detail that virtually tarnishes a product. New benchmarks will be needed to establish the true impact on performance, and although there should be very few occasions when a game demands all four gigabytes of video completely and at full speed, the error is still Nvidia's responsibility, which seems to have a black eye.
