Nvidia's Kepler architecture has displayed excellent, sometimes record-breaking performance since its grand debut in March. The GeForce GTX 680 was the fastest single-GPU graphics card until AMD recaptured the mojo last week, and the GeForce GTX 690 is still the undisputed king of single-card solutions, thanks to its dual GPUs (and prodigious thousand-dollar price tag).
In spite of these successes, Kepler has seemed eerily reluctant to get its feet wet in the shallow waters of the low-end market. Until earlier this month, the cheapest Kepler card Nvidia could offer cash-strapped enthusiasts was the GeForce GTX 670—which, at $400, really isn't very cheap at all. Anything more affordable was derived from Nvidia's older Fermi architecture and manufactured using a 40-nm fab process. All the while, AMD acolytes have had a whole lineup of state-of-the-art, 28-nm Radeons to choose from, including the $109.99 Radeon HD 7750 that came out back in February.
It couldn't last...
...and it didn't. In early June, Kepler finally jumped in the kiddie pool aboard the GeForce GT 640, an offering situated firmly in budget territory. The card launched at $99 and currently retails in the $99.99-109.99 range, almost squarely opposite the Radeon HD 7750. Nvidia makes a couple other versions of the GT 640, too, but those are for PC vendors, and you won't find them listed at e-tailers like Newegg and Amazon.
In any case, Nvidia has given buyers a very real option if they want Kepler goodness minus the lofty price tags. That, in turn, raises an important question. How well does the GeForce GT 640 fare against other budget cards vying for supremacy around the hundred-dollar mark? By that we mean not just the Radeon HD 7750, but also Nvidia's own, previous-gen offering.
More simply put, is the GeForce GT 640 a new budget wonder, or are you better off spending your lone Benjamin on another card? Let's find out.
Here's the star of our show, sans heatsink. Note the tiny GPU. That's the GK107, which packs 1.3 billion transistors into a scant 118 mm² footprint using TSMC's 28-nm fab process. The chip is slightly smaller than the Cape Verde GPU that powers AMD's Radeon HD 7700 series. Cape Verde measures 123 mm² and plays host to 1.5 billion transistors, though some of those transistors are disabled on the 7750, which has two fewer compute units than the 7770.
The GK107 is also much smaller than Nvidia's 40-nm GF106 and GF116 graphics chips, which drive the GeForce GTS 450 and GeForce GTX 550 Ti, respectively. Both of those older parts host just under 1.2 billion transistors in a die area of about 240 mm². Chips with physically smaller dimensions can cost less to produce, so although the GTS 450 is priced in the same neighborhood as the GT 640 now, Nvidia may have more freedom to apply future price cuts to the latter. So far, however, the GT 640 seems to be staying put at around $99.
Peer inside the GK107 with an electron microscope, and you'll see lots of transistors and gates arranged in all kinds of crazy patterns. The diagram above provides an easier-to-parse overview of the chip's various bits and pieces.
The GK107 features a single graphics processing cluster containing dual SMX shader multiprocessors. For reference, the GK104 chip inside the GeForce GTX 680 has four GPCs and eight SMXs. On both chips, each GPC contains two SMX units and a raster engine capable of rasterizing one triangle per clock. Each SMX has 192 arithmetic logic units (ALUs) and texture units capable of filtering 16 texels per clock cycle. The GK107's lone GPC is backed by a single ROP partition capable of producing 16 pixels per clock. The chip also has dual 64-bit memory controllers that give the GeForce GT 640 a 128-bit path to its 2GB of DDR3 memory.
Hold on a minute—DDR3?
Yes, believe it or not, Nvidia has equipped this card with DDR3 RAM instead of the speedier GDDR5. It's pretty sluggish DDR3, too, with an effective transfer rate of only 1782 MT/s. A version of the GT 640 with GDDR5 RAM does exist, but it's one of those pesky cards reserved for PC vendors and not available for sale to the general public.
It doesn't take a profound understanding of GPU architectures to guess that DDR3 could needlessly hamstring the GeForce GT 640 compared to its GDDR5-toting rivals. The peak theoretical numbers below lend weight to that notion:
|GeForce GTS 450||13||25||25||0.7||783||98|
|GeForce GTX 550 Ti||22||29||29||0.7||900||98|
|GeForce GT 640||14||29||29||0.7||900||29|
|Radeon HD 7750||13||26||13||0.8||800||72|
While the GeForce GT 640 compares favorably to AMD's Radeon HD 7750 in terms of peak shader, texturing, and rasterization throughput, it falls considerably short when it comes to memory bandwidth, with only 29GB/s to the Radeon's 72GB/s. We'll gauge the real-world performance implications of that shortcoming in a minute, but it certainly doesn't bode well. Graphics cards today need plenty of memory bandwidth to juggle textures and frame data.
The GT 640 may have one minor trump card, and that's its tight power envelope. Nvidia rates the card for peak power draw of only 65W, while AMD says the Radeon can draw up to 75W. The GeForce GTS 450 and GTX 550 Ti aren't even in the same league, with respective TDPs of 106W and 116W.