GeForce GTX 660: Introducing GK106
Yesterday, there were five GeForce cards in the channel based on Nvidia's Kepler architecture. Four of them centered on the company's GK104 graphics processor. One, the GeForce GT 640, employs a derivative GPU called GK107. Now, we have a second board with GK107 as its foundation.
But the second card surfacing today features an all-new Kepler-based chip called GK106.
GK106 is composed of five SMX clusters, each with 192 CUDA cores totaling 960. Sixteen texture units per SMX give us a total of 80 across the GPU. And three ROP clusters able to process eight full-color pixels yield 24 per clock cycle. A trio of 64-bit memory controllers aggregate into a 192-bit memory interface.
Nvidia claims that its GK106 is a fully-utilized GPU, and that there aren't any disabled resources we might see flipped on later. We find this odd, given that one SMB cluster standing on its own to the side.
Like the GeForce GTX 660 Ti, the reference 660 includes 2 GB of memory. The only way Nvidia is able to achieve this using three 64-bit memory controllers is by mixing capacities. Also like the 660 Ti, the only way this amount of memory can be handled over a 192-bit bus is with mixed-density ICs. It works like this: the three 64-bit controllers divide the total memory into 512 MB chunks, which are accessed at the full 192 bits. The remaining 512 MB is addressed by just one 64-bit controller in a completely separate transaction. Nvidia won't divulge anything else about its implementation for competitive reasons, but there is undoubtedly latency there the controllers have to contend with.
This is something we drill down into using bandwidth-specific benchmarks, and you'll want to keep it in mind in games where that extra 512 MB is utilized.
|GeForce GTX 660 Ti||GeForce GTX 660||GeForce GTX 570||Radeon HD 7850||Radeon HD 7870|
|Fabrication Process||28 nm||28 nm||40 nm||28 nm||28 nm|
|Core (Shader)/Boost Clock||915/980 MHz||980/1033 MHz||732 (1464) MHz||860 MHz||1000 MHz|
|Memory Clock||1502 MHz GDDR5||1502 MHz GDDR5||950 MHz GDDR5||1200 MHz GDDR5||1200 MHz GDDR5|
|Memory Bandwidth||144.2 GB/s||144.2 GB/s||152 GB/s||153.6 GB/s||153.6 GB/s|
|Graphics RAM||2 GB GDDR5||2 GB GDDR5||1.25 GB GDDR5||2 GB GDDR5||2 GB GDDR5|
|Power Connectors||2 x 6-pin||1 x 6-pin||2 x 6-pin||1 x 6-pin||2 x 6-pin|
|Maximum Thermal Design Power||150 W||140 W||219 W||130 W||175 W|
Given identical memory specifications, the GeForce GTX 660 Ti and 660 offer the same 144.2 GB/s of memory bandwidth. Most of the 660's other specs are less aggressive, though: CUDA cores, texture units, and ROPs are cut by roughly 30% in comparison. Nvidia helps overcome some of the impact of fewer resources by juicing the core and average GPU Boost frequencies, though.
Reference GeForce GTX 660
At first glance, Nvidia's reference GeForce GTX 660 reminds us of a GeForce GTX 460 or even the Radeon HD 7850. But then we turn it over...
...and the PCB only covers about 75% of the card's length. The overhang comes from the fan and shroud, suggesting that vendors can make much smaller versions of the GeForce GTX 660, providing they're able to address cooling effectively. Zotac sent us a sample of its card proving this is a feasible endeavor.
Again, Nvidia's reference card comes with 2 GB of GDDR5 at 1502 MHz. Its base core clock is 980 MHz, while GPU Boost averages 1033 MHz.
The reference GeForce GTX 660 exposes two dual-link DVI ports, calls for two DL-DVI, one HDMI port, and a DisplayPort output. This matches the GeForce GTX 660 Ti, and both cards support up to four monitors operating concurrently with three in Surround mode.
Nvidia's reference PCB measures 7" x 4.5", which is fairly small for a performance/mainstream card. One six-pin auxiliary power connector supplements the PCI Express slot's 75 W delivery, satisfying a TDP that stretches up to 140 W. A single SLI connector is tell-tale: the GeForce GTX 660 only supports dual-card configurations.