The gruesome details are here:

http://www.tomshardware.com/reviews/geforce-gt-640-review,3214.html

I gave away the punchline in the title, but I still need to back it up.  Nvidia's GK107 has a die size of 118 mm^2, while AMD's Cape Verde has a die size of 123 mm^2.  Both are on a TSMC 28 nm process node, though they're technically different process nodes.  But they're similar, so it's probably about the same cost per mm^2 to build the die.  Larger die sizes let you get more performance, but they're close enough that we should expect comparable performance, right?

Well, let's slant the comparison in favor of Nvidia.  Let's not use the top bin of Cape Verde, which is a Radeon HD 7770.  Let's use a Radeon HD 7750 instead.  And furthermore, let's ignore AMD's announcement that they're increasing the stock speeds on both the 7750 and 7770 by 100 MHz each.  And for good measure, let's ignore the fact that the Cape Verde cards have been available in volume since March, while GK107 still isn't available at retail.

So how does the new GeForce GT 640 fare?  It doesn't merely lose to the Radeon HD 7750.  It never even comes close, but rather, gets completely slaughtered in all seven games that the site tested.  I had expected the top GK107 card to be competitive with a Radeon HD 7750, as the latter gives about 1/4 of the performance of a GeForce GTX 680, while GK107 has 1/4 of the SMXes of a GTX 680 and 1/2 of the memory channels.

It gets worse for the GeForce GT 640.  It even has difficulty hanging with AMD's previous generation Radeon HD 6670, which is based on Turks.  Turks, like GK107, is also 118 mm^2.  But Turks is on an older 40 nm process node, which means that you can only fit about half as many transistors in the same die space.  The Radeon HD 6670 isn't an expensive card, either:

http://www.newegg.com/Product/Product.aspx?Item=N82E16814161386

Granted, that's a DDR3 version, rather than a GDDR5 version as in the Tom's Hardware review.  But look at the latter:

http://www.newegg.com/Product/Product.aspx?Item=N82E16814121442

$85 with free shipping before a $10 rebate.  And for a faster card than the $100+ GeForce GT 640.  This is the sort of result that makes me wonder if Tom's Hardware is doing something wrong in their review.

So how did the GT 640 end up so terrible?  For starters, no GDDR5 memory.  That's a mysterious choice for a card that should have been a competitor to a Cape Verde card that so obviously needs GDDR5.  Maybe there will be a future version with GDDR5 that isn't so embarrassingly bad, but it seems implausible that even that could catch a Radeon HD 7750.

Next, no GPU turbo.  I'm guessing that the die simply doesn't support it, though it might also be a case of crippling a die for the sake of market segmentation, as Intel does with their Core i3 processors.  But crippling your own die when you're getting slaughtered is terminally stupid.

There had been rumors that GPU turbo was kicking in more aggressively than Nvidia was letting on.  Overclocking a GeForce GTX 680 certainly produced far smaller gains than you'd expect from the clock speeds.  One French web site found that a retail GTX 680 that they bought was slower than the one Nvidia shipped them to review in every single game they tested, in spite of exactly the same clock speeds.  The difference was only about 2%, but that does feed rumors that some cards will have a more aggressive GPU turbo than others--and that the press got cherry-picked samples.

On another thread yesterday, I said that no GK106 meant that Nvidia had nothing to compete in the $150-$300 market.  Seeing how terrible the GeForce GT 640 is makes me wonder if I need to make that that the $100-$300 market.  And the GT 640 isn't competitive at its own price point at all, either.  That can be fixed by price cuts, of course--but judging from the last few generations, such price cuts may not be forthcoming.

The reason Nvidia couldn't compete at the low end the previous few generations was that their dies were too big, which meant their cards were too expensive to build.  I thought Nvidia had fixed that with Kepler, as GK104 is substantially smaller than Tahiti.  I expected GK107 to be a little slower than Cape Verde, but also to be compensate by being smaller.  It does far too much of the former and far too little of the latter to be competitive.

And it gets worse for Nvidia, too.  Above, I commented that AMD has said that they're increasing the stock clock speeds on the Radeon HD 7750 and 7770 by 100 MHz each.  So take a Radeon HD 7750 that is already far too fast for a GeForce GT 640 and add 10% or so to its performance by the time the GT 640 comes to market.

The sad thing is that people will still buy them for the brand. As each day goes by i have more and more trepidation for how the 660 is going to turn out. With their silence and chips like this releasing its hard to keep up any enthusiasm.

It would be a shame if the 670 ends up being the only decent thing they did this round as it doesnt bode well for us the consumers when it comes to prices.

I never really thought about a x40 nVidia product competing with a x700 AMD product.  When you look at the numbering its comparable product is much lower end.  Sure its the similair in die size and costs as much as the 7770, but the nVidia of the past was always like this.  It could just mean the 640 is a rebranding of an older architecture on a smaller process node.

• 690 - 7990
• 680 - 7970
• 670 - 7950
• 660 - 7870
• 650 - 7770
• 640 - 7670

---

Originally posted by Cleffy
I never really thought about a x40 nVidia product competing with a x700 AMD product.  When you look at the numbering its comparable product is much lower end.  Sure its the similair in die size and costs as much as the 7770, but the nVidia of the past was always like this.  It could just mean the 640 is a rebranding of an older architecture on a smaller process node.

• 690 - 7990
• 680 - 7970
• 670 - 7950
• 660 - 7870
• 650 - 7770
• 640 - 7670

---

There are multiple GeForce GT 640 cards.  One is a rebranded Fermi card.  The one that Tom's Hardware reviewed as linked above is a Kepler card, with GK107.  It has the Kepler feature set, with some stuff that was notably missing from Fermi, such as support for three monitors.

If GK107 isn't supposed to compete with Cape Verde, then what is?  GK106 seems to be completely missing, and how is that supposed to compete with both Cape Verde and Pitcairn?  Or is Nvidia going to have a bunch of severe salvage parts of GK104 to compete with Pitcairn?  That sounds like a losing cause, as GK104 is far more expensive than Pitcairn.

Also, a GK107 as slow as it is barely has any reason to exist.  It will probably be a little bit faster than Trinity integrated graphics, but not a lot.  How will it hold up against the integrated graphics in Kaveri or Haswell?  Why buy a discrete card that isn't much faster than integrated graphics?  If that's adequate performance in a new computer, then you'd just go with integrated graphics instead.  For people upgrading an older system, you can just use an older card instead, rather than using scarce 28 nm wafer starts on a card that doesn't make sense.