In a post on his company’s blog, nVIDIA CEO Jen-Hsun Huang addressed the controversy behind the last 512MB in the GeForce GTX 970’s 4GB VRAM. For those who do not know about this, the GTX 970’s memory is segmented into two portions: a 3.5GB segment, and another 512mb segment that operates at a vastly lower bandwidth. While few users play at high-enough resolutions to actually end up using that latter segment, analysis of performance indicates that there is indeed massive framerate stutter that occurs when going up to very high resolutions on this video card. The GTX 980, on the other hand, which also has 4GB of VRAM but has the same bandwidth all throughout, experiences only the expected performance loss associated with going up to higher resolutions, and nothing further.
Huang addressed this by telling people what they already knew: That this is actually exactly as nVIDIA intended the architecture of the GTX 970 to be, and that the reason why there’s so much of a controversy is that the technical marketing team wasn’t told of this little “feature,” which in turn resulted in reviewers not being informed of how this actually works.
That’s all well and good, but at the end of the day, even if the technical marketing team was informed about this, could there have been any reasonably-easy way to spin such an unorthodox memory architecture to the lay public? In this hypothetical alternate universe, we’d still have GTX 970s being sold with “4GB” plastered on boxes in big letters, and a tiny footnote saying that the last 0.5GB may experience memory loss. Are we really going back to the days when the amount of video memory was a marketing ploy targeted at hapless laymen? Clearly there was some misguided obsession at getting a 4GB sub-flagship off the factory floor, even if the latter portions of the memory meant that you’d get a bad experience anyway.
This is far from the first time that nVIDIA has had “marketing” issues. The infamous GeForce FX cards of old were proudly touted as having complete DirectX 9 support, but due to their choice of either inefficient FP32 precision or low-quality FP16 precision, suffered greatly compared to ATI’s Radeon 9000 series, which had support for FP24 and thus defeated the GeForce FX in performance in DirectX9 applications, while having comparable or better image quality. One generation before that, the GeForce 4 MX series was released to a public yearning for an upgrade, but instead found that the MX cards were weaker than the GeForce 2s and 3s that they were coming from!
Huang’s blog post comes off as just short of an apology without expressly stating it as such. With a class-action lawsuit at his doorstep, he’ll have to do better than just say “We’ll do better next time.” Perhaps rebrand the GTX 970s as 3.5GB cards with 512mb “caches.” Alternatively, albeit less realistically, rerelease the current GTX 970, rebrand it as 3.5GB, then create a “GTX 970 Ti” that uses binned GTX 980 parts and a proper 4GB memory pool.
Whatever the case, while this is a most unfortunate series of mishaps, it is more of a problem of company credibility and liability than a blow to the viability of the very competent GTX 970. Even those disappointed in nVIDIA still praise the GTX 970, and it remains a highly recommendable part with or without the controversial memory pool.