Another day, another set of rumors for Nvidia’s Maxwell GPU. This time however, it doesn’t concern the higher end Maxwell parts – but instead focuses squarely on the med to high range GTX 960 part. It’s been rumored for sometime the GTX 960 won’t be launching along with either the 980 or the 970, but the Japanese website Hermitage Akihabara are reporting that we’ll be seeing the GTX 960 in Mid October of this year. HA isn’t a website that usually prints bogus material, so while I’d not take the rumor as 100 percent accurate, it’s certainly very plausible.
There are two possibilities for the GTX 960 – the first is a rebrand of the current Kepler GPU. While this isn’t particularly likely – as it would go against the whole reason Nvidia skipped the 800 series – namely, confusion for customers, it is a slight possibility. If it is a re-branded Kepler then it’d make sense for it to sit between the GTX 760 and the GTX 770 in terms of raw performance.
The second and more likely scenario is the GPU will be a further cut down Maxwell GM204. If this turns out to be accurate then it makes logical sense for between 1408 and 1536 CUDA cores to be enabled on the device. We can gather this purely from the previous leaks concerning both the GTX 970 and 980.
A memory bus of 256-Bit is very likely – as anything more would be a waste, and a narrower bus wouldn’t provide the bandwidth required – unless the memory was clocked extremely high. It’s also a fairly good bet that we’ll see the part ship at 2GB. It’s possible we could see more, but it wouldn’t be super traditional for Nvidia to put in say 3GB. Board partners such as EVGA are of course free to release 4GB versions. Frankly if you’re planning on a pair of GTX 760’s for the purpose of SLI, this might be a better bet. 2GB is likely to start feeling limited for games in the next year or so, even if you’re only targeting 1080P.
Clearing up the die situation – currently industry speculation and ‘insider information’ predicts we’ll have version iterations of the Maxwell die. The first version will be at launch, and labeled as GM204-A1. This will be based on a 28nm process, since TSMC couldn’t perfect 20 or 22nm in time. Next, we’ll see GM204-A2. Supposedly this will release two to three months after the A1 revision. We’re not quite sure what this does, but there’s good chance it’ll be more stable clocks and overall be a nice improvement in yields.
Finally, we’ll get the GM204-B1 revision. The B1 is supposedly the big one (or should that be the small one), as Nvidia are shrinking down the die to the coveted 20nm process. We’re hearing there’ll be higher end Maxwell’s released at this point – such as the 980 Ti’s. But it’s also possible that we as customers won’t know Nvidia have done it – they’ll simply lower their costs. Another possibility as we’ll see 980+ or some other name of Maxwell. This simply indicates higher clock speeds.
This wouldn’t be the first time Nvidia have done similar – those who remember the 8000 GTX series will remember how Nvidia Milked it. The 320 / 640MB GTS, then a refined 8800 GT 512MB, then the 9800 series… made for a fairly boring time in the GPU industry for both reviewers and customers alike.
Either way, all we can do is wait and see what really comes out of all this. It will be very interesting to see how the mid range Maxwell’s end up overclocking, particularly given the reputation the GTX 750Ti managed to earn.