That Darn Memory Bus

Among the entire GTX 600 family, the GTX 660 Ti’s one unique feature is its memory controller layout. NVIDIA built GK104 with 4 memory controllers, each 64 bits wide, giving the entire GPU a combined memory bus width of 256 bits. These memory controllers are tied into the ROPs and L2 cache, with each controller forming part of a ROP partition containing 8 ROPs (or rather 1 ROP unit capable of processing 8 operations), 128KB of L2 cache, and the memory controller. To disable any of those things means taking out a whole ROP partition, which is exactly what NVIDIA has done.

The impact on the ROPs and the L2 cache is rather straightforward – render operation throughput is reduced by 25% and there’s 25% less L2 cache to store data in – but the loss of the memory controller is a much tougher concept to deal with. This goes for both NVIDIA on the design end and for consumers on the usage end.

256 is a nice power-of-two number. For video cards with power-of-two memory bus widths, it’s very easy to equip them with a similarly power-of-two memory capacity such as 1GB, 2GB, or 4GB of memory. For various minor technical reasons (mostly the sanity of the engineers), GPU manufacturers like sticking to power-of-two memory busses. And while this is by no means a true design constraint in video card manufacturing, there are ramifications for skipping from it.

The biggest consequence of deviating from a power-of-two memory bus is that under normal circumstances this leads to a card’s memory capacity not lining up with the bulk of the cards on the market. To use the GTX 500 series as an example, NVIDIA had 1.5GB of memory on the GTX 580 at a time when the common Radeon HD 5870 had 1GB, giving NVIDIA a 512MB advantage. Later on however the common Radeon HD 6970 had 2GB of memory, leaving NVIDIA behind by 512MB. This also had one additional consequence for NVIDIA: they needed 12 memory chips where AMD needed 8, which generally inflates the bill of materials more than the price of higher speed memory in a narrower design does. This ended up not being a problem for the GTX 580 since 1.5GB was still plenty of memory for 2010/2011 and the high pricetag could easily absorb the BoM hit, but this is not always the case.

Because NVIDIA has disabled a ROP partition on GK104 in order to make the GTX 660 Ti, they’re dropping from a power-of-two 256bit bus to an off-size 192bit bus. Under normal circumstances this means that they’d need to either reduce the amount of memory on the card from 2GB to 1.5GB, or double it to 3GB. The former is undesirable for competitive reasons (AMD has 2GB cards below the 660 Ti and 3GB cards above) not to mention the fact that 1.5GB is too small for a $300 card in 2012. The latter on the other hand incurs the BoM hit as NVIDIA moves from 8 memory chips to 12 memory chips, a scenario that the lower margin GTX 660 Ti can’t as easily absorb, not to mention how silly it would be for a GTX 680 to have less memory than a GTX 660 Ti.

Rather than take the usual route NVIDIA is going to take their own 3rd route: put 2GB of memory on the GTX 660 Ti anyhow. By putting more memory on one controller than the other two – in effect breaking the symmetry of the memory banks – NVIDIA can have 2GB of memory attached to a 192bit memory bus. This is a technique that NVIDIA has had available to them for quite some time, but it’s also something they rarely pull out and only use it when necessary.

We were first introduced to this technique with the GTX 550 Ti in 2011, which had a similarly large 192bit memory bus. By using a mix of 2Gb and 1Gb modules, NVIDIA could outfit the card with 1GB of memory rather than the 1.5GB/768MB that a 192bit memory bus would typically dictate.

For the GTX 660 Ti in 2012 NVIDIA is once again going to use their asymmetrical memory technique in order to outfit the GTX 660 Ti with 2GB of memory on a 192bit bus, but they’re going to be implementing it slightly differently. Whereas the GTX 550 Ti mixed memory chip density in order to get 1GB out of 6 chips, the GTX 660 Ti will mix up the number of chips attached to each controller in order to get 2GB out of 8 chips. Specifically, there will be 4 chips instead of 2 attached to one of the memory controllers, while the other controllers will continue to have 2 chips. By doing it in this manner, this allows NVIDIA to use the same Hynix 2Gb chips they already use in the rest of the GTX 600 series, with the only high-level difference being the width of the bus connecting them.

Of course at a low-level it’s more complex than that. In a symmetrical design with an equal amount of RAM on each controller it’s rather easy to interleave memory operations across all of the controllers, which maximizes performance of the memory subsystem as a whole. However complete interleaving requires that kind of a symmetrical design, which means it’s not quite suitable for use on NVIDIA’s asymmetrical memory designs. Instead NVIDIA must start playing tricks. And when tricks are involved, there’s always a downside.

The best case scenario is always going to be that the entire 192bit bus is in use by interleaving a memory operation across all 3 controllers, giving the card 144GB/sec of memory bandwidth (192bit * 6GHz / 8). But that can only be done at up to 1.5GB of memory; the final 512MB of memory is attached to a single memory controller. This invokes the worst case scenario, where only 1 64-bit memory controller is in use and thereby reducing memory bandwidth to a much more modest 48GB/sec.

How NVIDIA spreads out memory accesses will have a great deal of impact on when we hit these scenarios. In the past we’ve tried to divine how NVIDIA is accomplishing this, but even with the compute capability of CUDA memory appears to be too far abstracted for us to test any specific theories. And because NVIDIA is continuing to label the internal details of their memory bus a competitive advantage, they’re unwilling to share the details of its operation with us. Thus we’re largely dealing with a black box here, one where poking and prodding doesn’t produce much in the way of meaningful results.

As with the GTX 550 Ti, all we can really say at this time is that the performance we get in our benchmarks is the performance we get. Our best guess remains that NVIDIA is interleaving the lower 1.5GB of address while pushing the last 512MB of address space into the larger memory bank, but we don’t have any hard data to back it up. For most users this shouldn’t be a problem (especially since GK104 is so wishy-washy at compute), but it remains that there’s always a downside to an asymmetrical memory design. With any luck one day we’ll find that downside and be able to better understand the GTX 660 Ti’s performance in the process.

The GeForce GTX 660 Ti Review Meet The EVGA GeForce GTX 660 Ti Superclocked


View All Comments

  • CeriseCogburn - Thursday, August 23, 2012 - link

    ROFLMAO - the ONLY REASON you say you wanted the 7950 and it LOSES.

    There's the level of "your cred", you freaking loser.
  • Galidou - Saturday, August 18, 2012 - link

    This is only the point of the iceberg when we speak about credibility. Anandtech was nice enough to have a stock clocked part, we can't say that for most of the reviews on the internet.

    I even got on a website ''not gonna say it, could be too much shame for them'' that was comparing a non reference 660ti overclocked with... suspense... a 7850. And then some times in the review offered an ''alternative analysis'' against a 6870, who's dirty now?

    I won't name any but of all the review sites I usually read, they were all testing overclocked cards (plus the included Nvidia boost) against stock clocked AMD cards, ALL of them... Only one included minimum frame rates to all of the games tested which was interesting to see the limiting bandwidth acting at certain points. One can only wonder if the games released won't have any problem with that.

    I first came here on anand and almost pulled the trigger buying one RIGHT after finishing reading. Then I visited my other sites and it got all messed up. Anand didn't have minimum frames everywhere, others had different results, the games I play switch from one brand to another for the ''best bang for my bucks''.

    With all that mixed up mess, one can only wonder where the ''real'' truth is. I'll probably just end up buying a 7950 overclocking it 40-50% higher and not wonder about future games. At least I waited long enough to see the 660ti. Anyway the other reviewers had quite good result with the 7950 and it was STOCK omg 40-50% overclock can't give a bad performance...
  • CeriseCogburn - Sunday, August 19, 2012 - link

    *OC 660ti's on newegg and only 3 Stock.
    The author pointed out there is no default version, and Partners have a somewhat free reign on released clocks.
    Now be a good person and go look for yourself, you'll have a hard time finding a stock card vs an OC oob card.
    I'd also like to see that 40-50% 7950 OC....(methinks you really spewed overboard there)
    Reviews are noting a 17%-22% max performance gain on maximum 7950 OC, and that does not mean it's stable, except on a sole rider, non internet server, spanking clean, just defragged, built for benching, top of the line components, reviewer super massive rig.
    So, can we get that 50% OC bench set from you ?
    NO, of course we can't.
  • Galidou - Sunday, August 19, 2012 - link

    My friend bought the Twin frozr 3 while it was on special on newegg(300$ a week ago). overclocked 1150/1700 stable that's a 44% overclock and he could go higher, with the stock cooler. We reported gains of around 30 to 36% performance gain in games.

    On newegg, there's plenty of people reporting 1150 to 1200 core overclock, because it is in fact a 7970 board at a very cheap price. If you really can't accept one good thing about AMD that's where I differ from you.

    The thing is, Nvidia won this round for the average user, most of us don't overclock and are not fiddling with voltages and such. Including a nice boost is good for those average users, the fact is and whatever you might say, overclockers know it. AMD is very overclocker friendly this gen, end of the line, cry about it some more, it doesn't change the fact that they already know it, sorry. If you tried to misinform the people, you're too late, it's already circulating on the internet my friend.

    Now you shall say and I've heard it: ''People have been able to get their gtx680 overclocked to 1300 core in some cases so they are..........''. I know the drill, 680 has for the most part, a boost clock of around 1100 - 1150 boost clock. Lemme translate that, 200mhz overclock on a 1100 boost clock, 18% overclock on the cherry picked 680, because I'm comparing it with a 7950 which didn't pass the 7970 requirement.
  • CeriseCogburn - Thursday, August 23, 2012 - link

    Oh look at that, I didn't use a single fact again.
    you're pathetic.
  • CeriseCogburn - Thursday, August 23, 2012 - link

    the 660Ti's are hitting 1300+ cores.

    you're losing at stock out of the box in your highest triple monitor rez dummy

    Keep attacking like the fool you are.

    Now you may apologize profusely and thankme for saving you from your brainwashed amd embolism you claim to have acquired at overclock net
  • thebluephoenix - Monday, August 20, 2012 - link

    Cerise, as a punishment i would make you read few nvidia related articles at site called Semi Accurate to see why is so wrong to be biased idiotic crazy fanboy. Reply
  • CeriseCogburn - Thursday, August 23, 2012 - link

    Charlie is a liar, I am not. Deal with it. Reply
  • Galidou - Thursday, August 23, 2012 - link

    Everyone is a liar but you Cerise, all hail to you ohh great hardware god, I'm still waiting for news of you on you almighty owner of all the knowledge.

    Come and teach the nitrogen overclockers of the world about your so great knowledge about video card.
  • CeriseCogburn - Thursday, August 23, 2012 - link

    Yes, time for you to bow down, then thank me for having to correct you three times already, on the FACTS. Reply

Log in

Don't have an account? Sign up now