In January 2015, most of the tech news world was inundated with the news that the Geforce GTX 970, Nvidia’s best-selling mid-range GPU over any holiday period, was having issues with the on-board GDDR5 memory. As Redditors and people on various forums soon discovered, the last 512MB of VRAM on the card suffered from tremendously low throughput. Nvidia later revealed that this wasn’t an accident – a new design innovation in the Maxwell architecture allowed them to selectively disable faulty parts of a GPU in order to harvest more dies and improve overall yields. With Nvidia working on driver fixes for some of the games that misbehave on the card and trying to explain its way out of this mess, its CEO Jen-Hsun Huang stepped out of the shadows to address the controversy and offer Geforce owners an apology.
While some people moved to return their GPUs over this incident, and others line up to participate in a lawsuit against the company in the US, it didn’t seem to affect things too badly overall. Most people simply shrugged, claimed they were happy with performance regardless, and would wait for newer drivers to fix any memory issues that were affecting some games (Skyrim with mods being one of those). What happens inside the GTX 970 is the 256-bit memory bus is effectively cut up into a 192-bit segment and a 64-bit segment, or 3.5GB and 500MB VRAM pools respectively. Accessing that second pool of memory through the 64-bit bus forces overall speeds down to an eighth of the card’s possible performance. The L2 cache also takes a lop off, losing 256KB to end up with 1.75MB of on-board L2 cache instead of the full-fat 2MB that reviewers were told was on-board.
In addition, while Nvidia claimed in their reviewer’s guides that there were 64 ROPs available on the GTX 970, only 56 of those are actually in active duty on the card, while the remaining eight are sometimes used by the driver for anti-aliasing techniques like TXAA. It is on this basis, and the lack of L2 cache, that a lawsuit is being filed against Nvidia in the US Supreme Court, but I’m not sure if it’ll see any success or not. Only lawyers really benefit from lawsuits such as these.
In a blog post on Nvidia’s website, Jen-Hsun apologised for the fiasco and explains again that Maxwell’s handling of the separate memory pools is by design, not by fault. It appears to be a very sincere apology and I hope this will be the final nail in the coffin to this issue so that we may bury it. At the end, he puts it very bluntly, “We’ll do better next time.”
Some of you are disappointed that we didn’t clearly describe the segmented memory of GeForce GTX 970 when we launched it. I can see why, so let me address it.
We invented a new memory architecture in Maxwell. This new capability was created so that reduced-configurations of Maxwell can have a larger framebuffer – i.e., so that GTX 970 is not limited to 3GB, and can have an additional 1GB.
GTX 970 is a 4GB card. However, the upper 512MB of the additional 1GB is segmented and has reduced bandwidth. This is a good design because we were able to add an additional 1GB for GTX 970 and our software engineers can keep less frequently used data in the 512MB segment.
Unfortunately, we failed to communicate this internally to our marketing team, and externally to reviewers at launch.
Since then, Jonah Alben, our senior vice president of hardware engineering, provided a technical description of the design, which was captured well by several editors. Here’s one example from The Tech Report.
Instead of being excited that we invented a way to increase memory of the GTX 970 from 3GB to 4GB, some were disappointed that we didn’t better describe the segmented nature of the architecture for that last 1GB of memory.
This is understandable. But, let me be clear: Our only intention was to create the best GPU for you. We wanted GTX 970 to have 4GB of memory, as games are using more memory than ever.
The 4GB of memory on GTX 970 is used and useful to achieve the performance you are enjoying. And as ever, our engineers will continue to enhance game performance that you can regularly download using GeForce Experience.
This new feature of Maxwell should have been clearly detailed from the beginning.
We won’t let this happen again. We’ll do a better job next time.