NVIDIA GeForce GTX 660 Ti Review
In the months since its release, NVIDIA’s Kepler has reached impressive levels of popularity. The GTX 680 had waiting lists as long as we’ve ever seen, the GTX 670 remains a price / performance leader and the astronomically expensive GTX 690 is the fastest single card on the market. Unfortunately, there really hasn’t been anything geared towards sub-$350 price points but that’s all about to change with the GTX 660 Ti 2GB.
The GTX 660 Ti continues NVIDIA’s march down-market into lower pricing segments by targeting those with GTX 460 and HD 6870 series products that are looking for a significant performance boost in today’s more demanding games. Traditionally, the cards within the $199 to $349 price brackets are hot commodities and the GTX 660 Ti looks to be no different. This sweet spot segment has housed cards like the GTX 560 Ti, GTX 460, and now that their prices have been cut, the HD 7870 and HD 7950 as well. For those of you who need a quick recounting of history, every one of those products defined its generation and are still sought-after commodities.
So what differentiates a GTX 660 Ti from its more expensive and more capable siblings? The answer to that is twofold: at first glance not much has changed but the few differences which were made will have profound impact upon overall performance. Simply put, this card still uses the 3.54 billion transistor GK104 core we’ve come to know from other Kepler based cards but reduces capabilities in certain key areas.
Moving to this version of the GK104 seems to have been relatively simple and NVIDIA’s engineers deserve some credit for designing what seems to be a very scalable architecture. Instead of creating a whole new core with some base elements shared with the flagship parts -the GF114 and GF104 used this philosophy to great effect- NVIDIA simply optimized the existing GK104 by cutting back on the memory controllers, ROPs and L2 cache.
Since the memory controllers are tied at the hip to L2 cache and the ROP arrays, all three had to be cut back in order to achieve a balanced, adaptable solution for lower price brackets. Naturally, this will scale back performance when compared against a GTX 670 but these changes have also led to a significantly lower power and heat envelope than other Kepler-based cards.
The end result of NVIDIA’s cutting is a core with 1344 cores and 112 texture units spread across seven SMX engines with an operating frequency of up to 980MHz with GPU Boost, mirroring the specifications of a GTX 670. The similarities stop there since the GTX 660Ti will come with 24 ROPs and a 192-bit memory interface in order to distinguish it from other Kepler-based products. In plain English this means only 1/8th of the cores have been disabled but 1/4 of the ROP / Cache pipeline and one of the four memory controllers will be unavailable for the remaining cores. Will this cause a bottleneck? With the GTX 660 Ti’s operating frequency being equal to that of the GTX 670, likely not all that much but the performance hit in certain scenarios could be significant.
To stay competitive from a memory bandwidth standpoint, NVIDIA needed to use 2GB of GDDR5 but the 192-bit interface presented a challenge. If a standard layout would have been implemented, each of the three available controllers would have been populated by 512MB of memory, resulting in a somewhat paltry 1.5GB. Instead of going this route, the GTX 660 Ti uses a technology pioneered by NVIDIA’s GF116 and by extension the GTX 550 Ti: mixed memory allotments. This allows each channel to be populated by a different memory density and through the use of proprietary dynamic load balancing, the three memory controllers on the GTX 660 Ti’s core can adapt their throughout according to the architecture’s needs. In this case, NVIDIA has equipped two controllers with the standard 2x 256MB allotment while the remaining controller receives a quartet of 256MB modules.
At first, the 192-bit spec may be a turn-off for some in comparison to the HD 7870’s 256-bit layout but it has been paired up with the same 6Gbps GDDR5 modules found on the higher end card, partially mitigating the bandwidth loss from a narrow memory interface. Regardless of the inclusion of ultra gas GDDR5 memory, the GTX 660 Ti’s bandwidth is about 10% less than a HD 7870 so you can clearly see the importance of that extra 64-bit memory controller.
Pricing is an extremely important aspect of any card with mass gaming market ambitions and the GTX 660 Ti continues the $100 step-down trend set by NVIDIA and AMD in this generation. As solution that supposedly bridges a narrow gap between the HD 7870 and HD 7950, it receives an SRP of $299 which may force AMD to lower their prices again. This is slightly above $249 the launch price of NVIDIA’s own GTX 560 Ti but it lines up nearly perfectly with the often forgotten GTX 560 Ti 448 so the 660 should still present a great price / performance solution for gamers on a tighter budget. NVIDIA’s board partners will also be introducing pre-overclocked, custom designed SKUs with upgraded cooling solutions right alongside reference designs. These upgraded models should run in price from $309 up to $329 and above depending upon features and clock speeds. However, for the purposes of this review, we will be looking at a reference clocked model.
While some feel they have been waiting overly long for the GTX 660 Ti to hit the market, the timing of its release is perfect for the all-important back to school buying season. We’ve also been told by numerous board partners that stock will be ready for purchase on launch day and barring any massive upsurge in demand, there should be more than enough cards to go around. So for anyone that’s been on the sidelines, hoping NVIDIA’s capable Kepler architecture to reach more affordable levels, this could be the day you’ve been waiting for.
|Latest Reviews in Video Cards|