Welcome Stranger to OCC!Login | Register

NVIDIA Fully Details the Memory Configuration of the GeForce GTX 970

Category: Video Cards
Posted: 10:49PM

Recently some concerns have appeared over the NVIDIA GeForce GTX 970 and its 4GB of memory, namely that it wasn't using all of it. If a game needed to use all 4GB of memory, some people were not seeing the level of performance expected with the GTX 970. According to the specifications released when the GTX 970 launched, it has the same number of ROPs and L2 cache as the GTX 980. Unfortunately, that was never the case, as NVIDIA has now stated the GTX 970 has one fewer ROP and thus one fewer L2 cache to access. So the GTX 970 actually has seven ROPs and 1.75MB of L2 cache to play with, but still the same 1664 CUDA cores and 104 texture units.

So, NVIDIA has now established the ROPs and L2 cache is a little less than initially thought, and attributes that to a snafu between the engineering and PR teams. But what does that mean for the performance and total memory space? Well the GTX 970 may only have seven ROPs, but that 32-bit memory controller segment remains. In a way to counter that, NVIDIA made 3.5GB of memory available to seven memory controllers and their L2 cache, with 0.5GB to that final memory controller without its L2 cache. Many games would only need to access 3.5GB of memory, so there would not be an issue. However, if all 4GB needs to be accessed, then that final 0.5GB segment works at 1/7th the speed of the rest (yet still four times faster than system memory and PCI Express bus can handle). It may sound alarming, but that is only a 4-6% drop in real-world performance, according to NVIDIA's tests. The 224GB/s memory bandwidth, however, is no longer entirely accurate, as it is only met when all 4GB is used and not when it's just the 3.5GB section.

The miscommunication at launch between engineering, PR, and reviewers did result in places labeling the GTX 970 with an incorrect ROP/L2 count is unfortunate, and yes it did take four months for clarification from NVIDIA. However, the GTX 970 still can't technically be called a 3.5GB card, because it does actually have 4GB on it. Just maybe it should be broken down to 3.5GB with a 0.5GB cache. There are still eight memory controllers after all, and the GTX 970 runs on a 256-bit bus. If NVIDIA was still on Kepler and not Maxwell, we'd have a card with a 192-bit bus and only 3GB of memory. As it is, we get a card that didn't need to be hampered in that way, just one that wasn't entirely accurate in what it could do.

All in all, what we have with the GTX 970 is exactly what we should have. When games are accessing 3.5GB of memory or less, the performance is precisely how it should be in any review you can find on OCC or its affiliates. When all 4GB needs to be accessed, there is a small drop in performance, but it should not cause you or anyone else to abandon their GTX 970s for a 980 or equivalent card. NVIDIA made a mistake, has addressed the issue, and is making the consumer aware of what exactly they are getting. Should it have happened sooner? Yes, but we still have an incredibly powerful card that does exactly what it was always meant to. Just not always with accurate specifications. What that means to consumers on the whole remains to be seen, but I imagine any potential backlash (if any) won't be severe.

Source: PC Perspective

Register as a member to subscribe comments.

© 2001-2018 Overclockers Club ® Privacy Policy
Elapsed: 0.1629450321   (xlweb1)