Nvidia denies GeForce GTX 970 design flaw claims

Nvidia has denied that its GeForce GTX 970 graphics card design has a serious flaw, while admitting that performance can be marginally affected if more than 3.5GB of VRAM is in use at any one time.

Launched late last year, the GeForce GTX 970 boasts the company's new low-power Maxwell chip architecture on a board with 4GB of GDDR5 video memory. Where it differs from its higher-end GeForce GTX 980 sibling is in the number of unlocked portions of the GPU available to the user: while both use a GM204 GPU, the 980 features 4 GPCs and 16 SMMs to the GTX 970's 13 SMMs. That reduction in tessellation units corresponds to a clear drop in performance, made more obvious by a reduction in clock speed to further drop the power draw from 165W to 145W, but one which users on a budget or with thermal constraints will consider a fair trade-off.

Unfortunately, reports have been spreading since launch that there's a bigger problem in the design. Numerous customer support threads, both on Nvidia's own website and on retailer and enthusiast forums, have declared the design a bust thanks to an issue which makes it near-impossible to make full use of that 4GB of VRAM - a problem not shared by the GTX 980. According to these reports, detailed in full on Nvidia's official forum, when an application makes use of more than 3.5GB of the 4GB VRAM, performance is dramatically impacted.

Nvidia has been under pressure, both from the media and from its users, to declare its stance on the matter. Now, it has done so with a statement which admits to a 3.5GB VRAM segmentation configuration but denies that it causes an appreciable performance hit in gaming applications. 'The GeForce GTX 970 is equipped with 4GB of dedicated graphics memory,' an Nvidia spokesperson explained in a statement to press this week. 'However the 970 has a different configuration of SMs than the 980, and fewer crossbar resources to the memory system. To optimally manage memory traffic in this configuration, we segment graphics memory into a 3.5GB section and a 0.5GB section.'

That segmentation, Nvidia has admitted, can cause confusion as '3rd party applications that measure memory usage will report 3.5GB of memory in use on GTX 970' when an application requires less than or equal to 3.5GB of VRAM per draw. 'The GPU has higher priority access to the 3.5GB section,' the statement further explains, but denies that this 'higher priority access' has an appreciable effect on performance when a game requires more than 3.5GB. To illustrate its point, the company has performed internal testing between a GTX 980 and a GTX 970, running games at settings which require both more than and less than 3.5GB of VRAM.

The results: Shadows of Mordor shows a one percentage point difference in framerate between the GTX 970 and 980 results at >3.5GB settings; Battlefield 4 shows a three percentage point difference; Call of Duty: Advanced Warfare shows a similar three percentage point difference. 'As you can see,' Nvidia claimed, 'there is very little change in the performance of the GTX 970 relative to GTX 980 on these games when it is using the 0.5GB segment.'

This doesn't seem to square up with users' experiences, however, and we'd be surprised if GTX 970 owners who run at settings which require more than 3.5GB of RAM will be satisfied with the company's internal testing results - and even more surprised if they don't make a fuss that Nvidia failed to explain that a chunk of the card's VRAM was 'low-priority' and would be accessible only at the cost of performance, regardless of how small that cost might be.