Nvidia has been steadily extending its graphics card series based on 28-nanometer GK104 Kepler GPUs. The late-March release of the highly successful GeForce GTX 680 was followed by the early-May debut of the dual-processor flagship GeForce GTX 690 and, later on, by the GeForce GTX 670 which doesn’t yet seem to be mainstream to us, its recommended price being as high as $399. We do hope it will eventually become a mainstream product after a price cut. Right now, the new GeForce GTX 670 competes with the AMD Radeon HD 7950 which has the same recommended price. The competition is going to be exciting, especially as the GTX 670 seems to be even more attractive than the GTX 680…
Technical Specifications and Recommended Pricing
Technical specifications of the new Nvidia GeForce GTX 670 are given in the table below side by side with those of Nvidia GeForce GTX 680 (the differences are in bold font):
Design and Functionality
We’ve got a reference sample of the new card to test and it is hardly impressive. The face side of the PCB is covered by the black plastic casing of the cooler.
But when we take a look at the reverse side, we see an original solution: the short PCB is extended by means of a piece of plastic the cooler’s fan is fastened on. The GeForce GTX 670 is 241 millimeters long, which is 13 millimeters shorter than the GeForce GTX 680, but its PCB is a mere 170 millimeters long.
Like the reference GTX 680, the GeForce GTX 670 is equipped with two dual-link DVI outputs, one HDMI 1.4a connector and one DisplayPort version 1.2:
Thus, the GeForce GTX 670 can work with multi-monitor configurations (up to 4 displays simultaneously).
There are two 6-pin power connectors at the very edge of the short PCB.
The card’s peak power consumption is 170 watts in 3D mode as opposed to the GeForce GTX 680’s 195 watts. Using the MIO connectors located in their usual place, you can combine up to four GeForce GTX 670s into a SLI configuration.
The PCB seems to be kind of unfinished due to its small size. It looks as if it were sawn off along the memory chips.
So, the power circuit components are located in between the GPU and the video interfaces, just like on AMD’s Radeon HD 7870 and HD 7850 cards.
Managed by an ON Semiconductor NCP5392P controller, there are four power phases for the GPU and two for the graphics memory. That's a cheap and simple solution but how reliable is it?
The 28nm GPU of our card was manufactured in Taiwan on the 3rd week of 2012. It is revision A2 (the only revision spotted in retail so far).
As opposed to the full-featured GK104 Kepler, the GeForce GTX 670 is equipped with a simplified version of the GPU that has one out of the eight available SMX units turned off on the hardware level. The resulting GPU has 1344 unified shader processors and 112 texture-mapping units instead of the original version’s 1536 and 128. The number of raster operators has remained the same at 32. The base GPU clock rate in 3D mode is lowered from 1008 to 915 MHz at 1.175 volts while the GPU Boost technology can increase it by 65 MHz up to 980 MHz. The clock rate is dropped to 324 MHz at 0.987 volts in 2D mode.
The GeForce GTX 670 comes with 2 gigabytes of FCBGA-packaged GDDR5 memory in Hynix Semiconductor chips labeled as H5GQ2H24AFR R0C. They have a rated clock rate of 6000 MHz.
The 3D clock rate of the chips is 6008 MHz. Coupled with a 256-bit bus, this translates into a peak bandwidth of 192.3 GB/s, just like on the GeForce GTX 680. The memory clock rate is dropped to 648 MHz in 2D mode. The memory chips are located on both sides of the PCB and there are places for more chips there. So, we won’t be surprised to see 4GB versions of GeForce GTX 670 in the future.
Here is a summary of the reference Nvidia GeForce GTX 670 specs: