Over the weekend we reported on a memory allocation issue with Nvidia's GeForce GTX 970 graphics card. Some users were having difficulty using all of the card's 4 GB of VRAM, while others experienced performance issues such as stuttering when VRAM exceeded 3.5 GB.
Nvidia acknowledged the partitioning of the card's VRAM into a main 3.5 GB segment and a secondary 0.5 GB section might have been the cause. They also claimed that, in their testing at least, there was no performance reduction when VRAM use exceeded 3.5 GB, relative to the GTX 980 that has a different memory structure.
Now, Nvidia has made the interesting move of correcting their previously published specifications for the GTX 970. As AnandTech notes in their excellent breakdown on the matter, Nvidia incorrectly stated that the GTX 970 had 64 ROPs and 2 MB of L2 cache, when in fact the card has 56 ROPs and 1.75 MB of L2 cache.
This is because the GM204 GPU in the GTX 970 has one (out of four) ROP/memory controller partitions partially disabled. The ability to disable this particular section of the GPU is new to Maxwell's architecture, and wasn't picked up by Nvidia's technical PR team. It also explains why the GTX 970 has a different memory system to the GTX 980.
The correction to the GTX 970's specifications doesn't change its performance, which we found to be excellent for the price in our review last September. With proper caching of frequently accessed assets in the larger, 3.5 GB VRAM partition, it's unlikely that you'll notice issues relating to the slow memory performance of the remaining 0.5 GB.