14 DAY TRIAL // Over the past week, it came to light that the GeForce GTX 970 graphics card from NVIDIA, powered by the GM204-300 graphics processing unit, had trouble correctly allocating the memory installed on it.
This was not a “simple” matter of 500 MB of the 4 GB being unaccessible. The issue was that when a game tried to use more than 3.5 GB, performance dropped inexplicably and frame rates took a hit, causing lag.
More recently, the Santa Clara, California-based company responded to the controversy by saying that it's all got to do with how the streaming multiprocessors (SMs) in the GPUs don't work normally.
They had to split the 4 GB into 3.5 GB and 0.5 GB, with the latter having less than optimal memory access.
Now it appears that that was not the whole truth. The problem appears to be rather more serious than originally expected: the specs provided by NVIDIA back in 2014 were false.
The GeForce GTX 970 is weaker than stated
Back in September 2014, when NVIDIA first officially launched the GeForce GTX 980 and GTX 970, it said that the GM204-300 GPU of the latter had 1,664 CUDA cores, 104 TMUs and 64 ROPs.
Now, the company has revealed through PCPer that the information is partially false, due to a mix of an error in the reviewer’s guide and a misunderstanding between the engineering team and the technical PR team on how the architecture itself functioned.
The GM204-300 GPU only has 56 ROPs. Also, its cache memory isn't 2 MB like on the GM204-400 (GTX 980), but 1.7 MB.
Technically, the issue isn't the lower ROP count, since 56 raster operating units still output 56 pixels/clock. The bottleneck is in the SMMs.
The diagram below shows that the last Crossbar port has two memory blocks attached to it, which would cause twice the normal amount of requests under normal circumstances and, thus, making it stutter under heavy use.
To solve this, NVIDIA partitioned off the 0.5 block, leaving it to the OS to use it, which resulted in the 0.5 GB block to have a seventh of the speed of the 3.5 GB block.
This still makes it faster than memory over PCI Express by a factor of 4, but it seems that the clogging issue can still turn up under certain conditions when those 500 MB are accessed.
Does this mean GTX 970 is a 3.5 GB or 4GB card?
Technically, it's a 4 GB model, but we suppose a case could be made for it not being. Ultimately, it depends on whether or not you're one of the people who have issues with it. In the end, it's still a very powerful video adapter.
