We received our first GTX 970, the EXOC Black Edition from GALAX, at the very end of December, and we have been testing it against a recently purchased reference version of the 290X from VisionTek, as well as against our original non-throttling mildly overclocked PowerColor 290X PCS+. We have had a couple of extra weeks with our GALAX GTX 970 EXOC, which is more than usual as we attempted to investigate its memory controller whose specifications have recently been downgraded by Nvidia.
We tested our GALAX GTX 970 EXOC at factory-clocked speed, stock, and further overclocked, against the R9 290X and against the GTX 980, with 28 games at 2560×1600 and at 1920×1080 with the demanding settings that gamers use. We wanted to especially see if we could find any practical issues with the GTX 970’s memory controller which was respecified with a lower number of ROPs, less Cache, and downgraded from having a full 4.0GB of VRAM, to having a 4GB framebuffer with 3.5GB of fast access VRAM, plus 0.5GB with much slower access.
The Corrected Specifications of the GTX 970 and Nvidia’s apology
Very little has actually changed regarding the GTX 970 specifications since it launched in September, and since its overstated specification issues were brought to light about three weeks ago. Like the GTX 980 and the R9 290X, the GTX 970 comes equipped with 4GB of fast GDDR5 video memory (vRAM). Unlike the GTX 980 and the 290X, the GTX 970 uses a memory controller which primarily addresses 3.5GB, and secondarily accesses 0.5GB of vRAM much more slowly because of shared resources.
Nvidia’s technical marketing team members claimed that they were originally unaware of their Maxwell architecture’s ability to individually disable ROPs. This meant that the actual ROP count, was originally misrepresented as 64 instead of 56, and the L2 cache was given as 2MB instead of the correct 1.75MB. Nvidia was quick to apologize and to give us a detailed look at the GTX 970’s memory controller.
The difference with the original overstated ROP count is because of limitations with the memory controller’s crossbar. The crossbar routes the data from the processing clusters to the memory controller using the L2 cache. By limiting the L2 cache to 1.75MB, Nvidia had to disable access to the GTX 970’s cache for one of the eight controllers by routing it to another memory controller’s cache which makes it slower for that controller. The original bandwidth specification of 224GB/s was properly lowered to 196 GB/s by Nvidia.
The memory is divided into two parts – one with 3.5GB of the vRAM composed of memory controllers with direct access to the L2 cache – and into a second 0.5GB part sharing access to the cache with another memory controller. A game can only access one section of the vRAM at a time – either it will access the 3.5GB block or the 0.5GB block, but not simultaneously. However, if both segments are being read simultaneously, there will be a noticeable decline in performance compared to the GTX 980’s memory controller which has no such limitations. However, Nvidia said that they designed the GTX 970 to work around these issues with drivers, and when these rare situations happened at extreme resolutions, the performance hit would only be 1-4% compared with the GTX 980’s controller.
When Nvidia published the original specifications, they neglected to give these important details to the reviewers as they have done in the past with other hybrid GTX memory controllers. Technically, Nvidia was correct in stating that the GTX 970 ships with 4GB of vRAM, but they did not give full disclosure.
The issues with “false advertising” primarily rose when Nvidia misstated the number of ROPs as 64 when there are actually 56. However, this technical misstatement does not affect the bandwidth nor the performance of the GTX 970 as it is unable to utilize the full 64 ROPs anyway, due to the limited amount of pixels the shaders send to the ROPs per clock cycle. The GTX 970 is already bottlenecked when it comes to fill rates, unlike with the GTX 980, because for basic pixel operations the GTX 970 is primarily bound by its lower SMM count.
This diagram that Nvidia provided to us and to the other Tech Editors last week shows the correct way in which the GTX 970 accesses its memory.
The L2 cache was originally misstated as 2MB but it is actually 1.75MB, and it has been corrected along with number of the ROPs, the bandwidth figures have been downgraded, and the memory controller has been explained in the updated technical specifications, and the GTX 970 Reviewer’s Guide has also been revised.
The Forum Angst and the AMD Advocates
AMD’s army of Advocates certainly have seized on this issue, roundly condemning Nvidia for what their own company has practiced in the past. In one case, AMD overstated the transistor count for Bulldozer’s FX-8150 by 800 million although it actually has only 1.2 billion and not 2 billion as was originally published months earlier; and in another, when they told the reviewers the week of its launch that the HD 6990’s overclock BIOS was covered under warranty, and then they changed their minds a few days after the reviews were published.
Nvidia admitted to their “miscommunication” but insists that the GTX 970 is still the same card with what they call incredible performance for the price, and they have even assisted dissatisfied buyers of the GTX 970 to return, refund, and/or upgrade their card even if the upgrade period normally allowed by resellers had expired. Yet, three weeks later, the tech forums are still buzzing with tech posters posting the same inflammatory and negative comments about Nvidia’s “deceit”.
The Nvidia GeForce forums have exploded with angry comments all over the world. In the USA, there are three hundred pages on this single topic that would make it seem to be a PR catastrophe for Nvidia. Yet the worldwide petition to force Nvidia to refund money to GTX 970 buyers has no traction. Less than 6,000 people signed – out of hundreds of thousands of GTX 970 owners!
Looking closely at the anti-GeForce threads, one thing becomes very clear – you will see the same pictures, the same names, and the same users posting their anger over and over and over again – basically the same posters posting in multiple threads across many tech forums. If they had any sense and are really legitimate owners of the GTX 970, one wonders why they haven’t refunded their GTX 970s by now if they are really as angry as they claim to be.
So where does this negativity and anti-Nvidia sentiment come from? One can only assume that AMD’s army of thousands of “Advocates” are at work. It appears that AMD’s “targeted” marketing to Social Media which began over 6 years ago under former CMO Nigel Dessau is largely a failure in influencing buying decisions although it has negatively impacted our tech communities.
Real World Gaming and the GTX 970
Does the GTX 970’s new limitations actually impact performance in games? According to Nvidia, the maximum performance hit for rare situations where the 0.5GB of memory is accessed in games is minimal – only up to about 4% even at 4K and with maximum details turned on. The GTX 970 is designed to split its memory buffer this way and Nvidia says the card is working exactly as intended..
We went looking for these problems with the GTX 970 versus the GTX 980 (and versus the R9 290X) in several ways. We tested the GTX 970 against the GTX 980 and attempted to find problematic games that would stress the two memory controllers simultaneously.
Besides testing with 28 games, we focused on benching several of the console ports that use a lot of video memory like Middle Earth: Shadows of Mordor, BF4 ,and Dragon’s Age; Inquisition. The GTX 970 may take a slightly larger hit in certain circumstances than the GTX 980 but when you are in situations where the 0.5 GB limited memory controller access begins to affect performance, you are already in unplayable situations for both cards.
Our conclusion: At 2560×1600, and at even higher 4K resolutions with DSR applied, we could not find a single situation where the memory controller affected the framerates where they were not already too low for game play. Perhaps GTX 970 SLI might have some issues, but that testing is beyond the scope of this evaluation.
What we now really want to see is the GTX 970’s performance – in the form of stock clocks, at GALAX EXOC clocks +76MHz offset over reference, and overclocked as far as we can go on stock voltage and fan profile. And we want to compare its performance with the GTX 980 and with its main competitor, the R9 290X, which is priced similarly to the GTX 970. Of course, we shall also give you our “Big Picture” to place the GTX 970 into a line up with 7 other AMD and Nvidia cards.
First, lets unbox the GALAX GTX 970 EXOC and look at it specifications and features: