These days news about the NVIDIA GeForce GTX 970 3.5GB of video memory are in the air. Yesterday, Xiao Bian has also verified this problem in lots of games, it shows only 3.5GB utilization. NVIDIA has also made explained that this is by design. But the official answer did not explain in detail and there are lots of questions in the consumer minds. Now the foregin media PCPer has dig this matter into deep and find out that, there is an error on the NVIDIA GTX 970 graphics card specifications as it is featuring only  56 ROPs and about 1.7MB of L2 cache available to it. It can still hit the bandwidth of 224 GB/s but in order to achieve these numbers the card has to access both the memory blocks. 

NVIDIA Geforce GTX 970  Specifications correction: ROP and L2 cache downsizing


NVIDIA GTX 970 specifications previously described in the official documentation are wrong, the card is used Maxwell GM204 core same featuring in GTX 980, NVIDIA just says it before SMM unit shielded three groups in order to reduce power consumption, 7GHz memory frequency, 256bit bit wide and 64 ROP units reserved, but in fact some of the specifications are wrong.

There must be differences in numbers behind the GTX 970 graphics memory with the exception and now NVIDIA has recognized the original GTX 970 specifications are wrong, 4 ROP / memory control partitions are actually disabled, so means it has only 56 ROP units and L2 cache less than 2MB about 1.75MB, but the memory controller does not change, indeed 256bit bit wide and 4GB of GDDR5 memory.

TX 970 three groups SMM of units are disabled
TX 970 three groups SMM of units is disabled

Partially disabled ROP / memory master function is a new feature on the Maxwell architecture only, this feature will be introduced later in the original (Maxwell Gm200).

As to why there is such a mistake, NVIDIA’s explained that it is accidental, such documentation has prepared by NVIDIA’s technical sales team. (Power Point Presentation was also a staff mistake?) NVIDIA explain that technical sales team in making the evaluation guidelines do not realize that Maxwell architecture is with “partial ban” feature mentioned above, they know there is a 256bit GTX 970-bit wide, but I do not know can independently prohibit ROP units and 2MB L2 cache, so they made an error in the documentation of various departments in the NVIDIA spread open, and through the media away.

This problem has not been found in the past four months, until the beginning of the investigation until this month, NVIDIA GTX 970 graphics memory addressing issues were noticed.

NVIDIA is now facing a public relations crisis undoubtedly. In addition Anandtech website has done a reflection, in fact, before these problems also exposed some clues, such as NVIDIA’s CUDA has been reported before the DeviceQuery tool GTX 970 graphics L2 cache is 1.75MB, while the GTX 980 is 2MB.

Some Tools have also been reported in the past before the L2 cache streamlined case of GTX 970
Some Tools have also been reported in the past before the L2 cache streamlined case of the GTX 970
GPU-Z sort of third-party software does not detect this problem
GPU-Z sort of third-party software does not detect this problem

GPU-Z, AIDA64 and other third-party softwares are not aware of the situation, so they cannot detect the problem, the latest version of GPU-Z 0.8.1 still shows a 64 ROP units.