NVIDIA GT200 Revealed – GeForce GTX 280 and GTX 260 Review

The Products – GeForce GTX 260 896MB Odd-ball and Naked GPU

Nope, that’s not a typo – the GeForce GTX 260 comes with 896MB of frame buffer; that’s 1024MB – 128MB for those keeping count at home.  This should turn on a light in our heads after having looked the GT200 architecture in such detail previously: the GTX 260 uses fewer shader processors, fewer ROPs and has a smaller memory bandwidth because of it.  Each ROP/memory controller references 128MB of memory and NVIDIA took one of the sets of ROPs/memory controller and removed them from the architecture to make this slightly slower, slightly less expensive model.  Of course, we more than likely seeing a binning/yield solution at play more than anything else – rather than throwing away cores that turned out to have one bad set of SPs or memory blocks they can just disable both and sell it as a GTX 260. 

The GeForce GTX 260 runs at a core clock rate of 576 MHz while its 192 shader processors (notice this is 48 less than the GTX 280) operate at 1242 MHz.  The 896MB of GDDR3 memory is running at 999 MHz – why they didn’t step up to 1000 MHz is beyond me, so don’t ask.  Our GTX 260 sample was a reference card from NVIDIA so these clocks are the defaults.  For this card, NVIDIA labels the board power maximum of 183 watts, and again we’ll see as our testing progresses how that pans out.

The pricing on the GTX 260 is much more reasonable, with an MSRP from NVIDIA of $399.

The card itself is virtually identical to that of the GTX 280:


It has the same dual slot cooler design, same surrounding chassis, same SLI connections and the only difference is in regards to power connections.



Unlike its bigger brother that requires a 6-pin and 8-pin connection, the GTX 260 only requires a 2×6-pin PCIe power configuration. 

Uncovering the Beast Within

Taking apart the GTX 280 or GTX 260 cards was quite a pain in the rear.  Besides having a cool dozen screws to remove, the two pieces are essentially clipped together at various points along the perimeter of the card.  My strategy for removing them was to essentially pull them apart until they flexed and then find that clip to pull it some more.  I’d highly recommend you take a much more careful approach with anything you actually purchase though…


Taking off the heatsink reveals this: a very beefy copper block surrounded by aluminum heatsinks for the memory, display chip and power controller logic.  The areas that look spray painted are, in fact, spray painted, so that the only color you see from the outside is black; I’d compare this to the practice of painting the joists under your deck black to attempt to “hide” them from view. 


Looking at the PCB under that heatsink we find the GT200 GPU, eight memory chips of 64MB capacity each, lots of digital MOSFETs, video output chip we keep referring to – and a nice shiny quarter!  Okay, I probably put that there for size comparison but it would have been a nice surprise.


Here is a closer view of the GPU with good ‘ol George Washington.


After my testing I decided to remove the cores from one of the GTX 280s and an old 8800 GTX so I could compare them side to side.  While the GT200 is definitely larger, it is not dramatically larger just by eying it.  This particular GT200 sample was built in the 3rd week of 2008 while the G80 core was built in the 24th week of 2006 – in case you were curious. 

(Note to readers that might have passed out: I was just kidding about ripping off any GPUs from cards.)