Last month, we detailed how Nvidia’s GTX 970 has a memory design that limits its access to the last 512MB of RAM on the card. Now the company is facing a class-action lawsuit alleging that it deliberately mislead consumers about the capabilities of the GTX 970 in multiple respects.
Nvidia has acknowledged that it “miscommunicated” the number of ROPS (Render Output units) and the L2 cache capacity of the GTX 970 (1792KB, not 2048KB), but insists that these issues were inadvertent and not a deliberate attempt to mislead customers. There’s probably some truth to this — Nvidia adopted a new approach to blocking off bad L2 blocks with Maxwell to allow the company to retain more performance. It’s possible that some of the technical ramifications of this approach weren’t properly communicated to the PR team, and thus never passed on to reviewers.
Nvidia’s decision to divide the GTX 970’s RAM into partitions is a logical extension of this die-saving mechanism, but it means that the GPU core has effective access to just seven of its eight memory controllers. Accessing that eighth controller has to flow through a crossbar, and is as much as 80% slower than the other accesses. Nvidia’s solution to this problem was to tell the GPU to use just 3.5GB of its available memory pool and to avoid the last 512MB whenever possible. In practice, this means that the GTX 970 flushes old data out of memory more aggressively than the GTX 980, which will fill its entire 4GB buffer.
Of principles, practices, and performance
Nvidia has maintained that the performance impact from disabling the last 512MB of memory in single-GPU configurations is quite small, at around 4%. our own performance tests found little reason to disagree with this at playable resolutions — at 4K resolutions we saw signs that the GTX 980 might be superior to the 970 — but the frame rates had already fallen to the point where the game wasn’t very playable. At the time, I theorized that SLI performance might take a hit where single-GPUs didn’t. Not only is there a certain amount of memory overhead associated with multi-GPU performance, two graphics cards can drive playable 4K resolutions where a single card chokes. We’ve been recommending that serious 4K gamers explore multi-GPU configurations and the GTX 970 appeared to be an ideal match for SLI when it first came out.
Testing by PC Perspective confirmed that in at least some cases, the GTX 970’s SLI performance does appear to take a larger-than expected hit compared to the GTX 980. Still, they note that the issues only manifest at the highest detail levels and graphics settings — the vast majority of users are simply unlikely to encounter them.
One thing that makes the complaint against Nvidia interesting is that the facts of the case aren’t really in dispute. Nvidia did miscommunicate the specifications of its products and it didmisrepresent those figures to the public (advertising 4GB of RAM when only 3.5GB is available in the majority of cases). The question is whether or not those failings were deliberate and if they resulted in significant harm to end users.
The vast majority of customers who bought a GTX 970 will not be materially impacted by the limits on the card’s performance — but people who bought a pair of them in SLI configurations may have a solid argument for how Nvidia’s failure to market the card properly led them to purchase the wrong product. It’s also fair to note that this issue could change competitive multi-GPU standings. AMD’s R9 290X starts $ 10 below the GTX 970 at NewEgg, but has none of the same memory limitations. The GTX 970 is still a potent card, even with these limitations and caveats, but it’s not quite as strong as it seemed on launch day — and there are obviously some Nvidia customers who feel mislead.
Cases like this are often settled before they ever go to trial, but Nvidia has yet to comment publicly on the lawsuit.