GK104 takes a step down
NVIDIA’s latest graphics card is a new GK104-based GTX 670 with a lower price tag that also competes against the HD 7950 and HD 7970.
While the graphics power found in the new GeForce GTX 690, the GeForce GTX 680 and even the Radeon HD 7970 are incredibly impressive, if we are really honest with ourselves the real meat of the GPU market buys options much lower than $999. Today’s not-so-well-kept-secret release of the GeForce GTX 670 attempts to bring the price to entry of the NVIDIA Kepler architecture down to a more attainable level while also resetting the performance per dollar metrics of the GPU world once again.
The GeForce GTX 670 is in fact a very close cousin to the GeForce GTX 680 with only a single SMX unit disabled and a more compelling $399 price tag.
The GTX 670 GPU – Nearly as fast as the GTX 680
The secret is out – GK104 finds its way onto a third graphics card in just two months – but in this iteration the hardware has been reduced slightly.
The GTX 670 block diagram we hacked together above is really just a GTX 680 diagram with a single SMX unit disabled. While the GTX 680 sported a total of 1536 CUDA cores broken up into eight 192 core SMX units, the new GTX 670 will include 1344 cores. This will also drop the texture units to 112 (from 128 on the GTX 680) though the ROP count stays at 32 thanks to the continued use of a 256-bit memory interface.
In fact, the memory system on the GTX 670 remains unchanged from the GTX 680 including a 2GB frame buffer running on the 256-bit bus at a reference speed of 6 Gbps totalling 192.2 GB/s of available memory bandwidth.
Reference clock speeds on the GTX 670 will be set at a base clock of 915 MHz and a Boost clock of 980 MHz – a delta of 65 MHz. That gap is a bit wider than what we saw on the original GTX 680 GPU of 58 MHz. Memory clocks will be operating at 1500 MHz or 6 Gbps at their stock settings.
The obvious benefit to this revision of the GK104 GPU is that NVIDIA can utilize parts that might have otherwise not been able to pass requirements to be labeled a full-speed GTX 680 and thus utilize die they have already paid for from TSMC. The GTX 570 was built with the same idea in mind, as was the GTX 470 and so on.
Galaxy GeForce GTX 670 2GB Graphics Card
Unlike both the GTX 680 and the GTX 690 launch, retailer cards will vary quite a bit on launch day as NVIDIA allowed their partners to have early access to the designs for the GTX 670 (hence the leaks obviously) and built their own custom solutions. Our testing today is based around a reference-clocked Galaxy GTX 670 2GB card as well as a reference card from NVDIIA – so we are going to show you base results. Of course we spend some time on overclocking later in the article.
The Galaxy card offers some familiar branding, but sadly, no "dispalyport" to be had…
The reference GTX 670 cards will share the same cooler design as the GTX 680 though there are some pretty substantial differences to the PCB and board design. Our Galaxy and reference cards are also MUCH lighter than the GTX 680 with a slightly less dense cooler required for this GPU.
Right away, something is different – the PCB doesn’t actually go to the end of the card. The fan itself actually extends past the PCB and the power connectors can be found right on the end of the circuit board. NVIDIA did this to demonstrate that the GTX 670 could be built in very small form factors (think the Alienware X51 system) or be expanded to full size coolers for better performance.
With a TDP of 170 watts, the GTX 670 still requires a pair of 6-pin power connectors to operate but don’t let the odd placement bother you.
The GTX 670 will support 3-Way
and 4-Way SLI as you can see from this pair of SLI connections!
Display configurations will vary more with each vendor’s custom cards, but the reference design includes a pair of dual-link DVI connections as well as a full-size HDMI and DisplayPort connection.