Generally, you don't see many products released in December. It's getting a little too late to make dramatic impacts on Q4 earnings as many have already done their holiday shopping. If a company is going to release a new product this late in the year, there's generally a good reason for it, or it's simply a product we'll never see.

At the end of October NVIDIA introduced its GeForce 8800 GT, based on a brand new 65nm GPU codenamed G92. The 8800 GT quickly outclassed virtually every NVIDIA GPU, making most of the G80 lineup obsolete by offering better performance at lower prices. A higher end incarnation of the 8800 GT's G92 was inevitable, we just didn't expect to see it this soon.

The GPU is the same, we're still looking at a G92 derivative part, but the card is all new: the GeForce 8800 GTS 512.


A dual-slot G92, the larger heatsink keeps this card a bit cooler than the 8800 GT but with no increase in sound

While NVIDIA is in a better position than AMD is these days, NV marketing could stand to learn from AMD's recent changes. The Radeon HD 3800 series carry no tacky suffixes, just four digit model numbers to keep things nice and simple. Not only is the GeForce 8800 GTS 512 absurdly long, it also further complicates the 8800 product line. If you'll remember back to our 8800 GT review, the 8800 GT is faster than the old G80 based 8800 GTS. The new 8800 GTS 512 is faster than the 8800 GT, and thus faster than both the 320MB and 640MB versions of the old GTS. So you end up with the following lineup today:

8800 Ultra > 8800 GTS 512 > 8800 GTX > 8800 GT > 8800 GTS 640 > 8800 GTS 320

Confusing to say the least, but if you can forget about all of the other products on the market you'll see that there are only two NVIDIA cards to be concerned with: the 8800 GTS 512 and the 8800 GT.


Form Factor 8800 Ultra 8800 GTX 8800 GTS 8800 GTS 512 8800 GT 256MB 8800 GT 8600 GTS
Stream Processors 128 128 96 128 112 112 32
Texture Address / Filtering 32 / 64 32 / 64 24 / 48 64 / 64 56 / 56 56 / 56 16 / 16
ROPs 24 24 20 16 16 16 8
Core Clock 612MHz 575MHz 500MHz 650MHz 600MHz+ 600MHz+ 675MHz
Shader Clock 1.5GHz 1.35GHz 1.2GHz 1.625GHz 1.5GHz+ 1.5GHz+ 1.45GHz
Memory Clock 1.8GHz 1.8GHz 1.6GHz 1.94GHz 1.4GHz - 1.6GHz 1.8GHz

2.0GHz

Memory Bus Width 384-bit 384-bit 320-bit 256-bit 256-bit 256-bit 128-bit
Frame Buffer 768MB 768MB 640MB / 320MB 512MB 256MB 512MB 256MB
Transistor Count 681M 681M 681M 754M 754M 754M 289M
Manufacturing Process TSMC 90nm TSMC 90nm TSMC 90nm TSMC 65nm TSMC 65nm TSMC 65nm TSMC 80nm
Price Point $600 - $800+ $500 - $600 $270 - $450 $349+ $219 - $229 $299 - $349 $140 - $199

 

Architecturally, the 8800 GTS 512 adds another group of 16 shader processors over the 8800 GT. We'd suspect that the 8800 GT has the same number of SPs, but with one block of 16 disabled to increase yields.

Since it's based on G92 we get a 1:1 ratio between texture address and texture filtering, giving the GTS 512 the first leg up over the much more expensive 8800 Ultra. With a 650MHz core clock and 1.625GHz shader clock, the GTS 512 has an 8% shader processing advantage over the Ultra.

The only area where the 8800 GTS 512 loses to the 8800 Ultra is in its total memory bandwidth. The 8800 Ultra, like the 8800 GTX, features a 384-bit wide memory bus while the GTS 512 uses the same 256-bit memory interface from the 8800 GT. There are definite cost advantages to going with a 256-bit memory bus; NVIDIA can build a smaller chip with fewer pins, and make up for the loss in memory bandwidth by shipping the card with faster memory devices. Despite the 1.94GHz memory data rate on the 8800 GTS 512, the 8800 Ultra and GTX have around a 40% memory bandwidth advantage, resulting in better performance in memory bandwidth limited scenarios and high resolution AA tests.

Despite being built on a 754M transistor die, the move to 65nm has made G92 much smaller and thus cheaper to make than G80, which is why we're seeing NVIDIA eagerly replacing its 8800 lineup with G92 variants.

Pricing and Availability

With the disappointing aftermath of the 8800 GT launch, we're better prepared to analyze expectations for what will happen with the 8800 GTS 512. Keep in mind that the 512MB 8800 GT is supposed to be a $250 part, but in reality it's selling for around $300 in the US. The GTS 512 is expected to sell for $299 - $349, but we're already hearing from manufacturers that prices will be much higher.

The XFX GeForce 8800 GTS 512 at reference clocks will carry an MSRP of $349, and the overclocked XXX edition will sell for $379. The GTS 512 could possibly sell at $349, but we wouldn't be too surprised to see it priced even higher in the market given its close proximity to the 8800 GT.

The 8800 GT 256MB: Here at Last
Comments Locked

56 Comments

View All Comments

  • Griswold - Thursday, December 13, 2007 - link

    Also (partly) wrong. Its a good price/performance part and its short in supply. That is why its priced higher. And I'm willing to bet the supply shortage is artificial. Look at how the availability of the GTS 512 is - seems to be much better than that of the GT. Its no surprise. Nvidias margins with the GT must be abyssmal compared to that of higher priced units (thats a given, but they also rendered almost their complete lineup obsolete for several weeks prior to the launch of the GTS 512), but they needed that horse to compete with the 3850/3870 price point.

    And you really need to stop talking out of your ass about the 3850. Its selling well and its selling at MSRP because supply is decent (and you lecture him about fundamentals... ). I think there was a the register claim of 150k units in 3 weeks. Well, thats three times the amount of the available 8800GT units in the same timeframe. Speaks for itself.
  • neogodless - Tuesday, December 11, 2007 - link

    Whew... just bought an 8800GT and would like to feel like it was a good buy for a *little while*! Hope it has enough supply to help drive prices down in general though...
  • R3MF - Tuesday, December 11, 2007 - link

    Where are the G92 GTS cards with memory over 2.0GHz?
    Does this preage the entrance of a G92 GTX with memory at 2.4GHz and a higher core clock?

    It isn't rocket science to put some decent speed memory on a midrange card. Witness the 3870 with 2.35GHz memory, so why haven't any of the so called "OC" versions of the G92 GTS got overclocked memory?

    At the same time we all want a card that can play Crysis at 1920x1200 at High details and still get around 30FPS. The GTS can get ~30FPS at Medium details........... whoopy-do!

    So, we know its possible to economically provide more bandwidth and we know its necessary, but nobody has done so including the OC'ed versions.

    Is this because there is a G92 GTX product around the corner?

    Yes i know there is rumoured to be a G92 GX2 dual-card sometime in january, but how about a non-cack single card version.

    A card with:
    720MHz core clock
    2000MHz shaders
    2400MHz memory
    1GB or memory

    would absolutely rock, so why haven't we got one?
  • kilkennycat - Tuesday, December 11, 2007 - link

    Memory tweaking of the current series is a tiny marginal benefit with a huge increase in power-dissipation. The G92 represents the last gasp of the current G8x/G9x architecture. The shrink was absolutely essential to nVidia's GPU business to get away from the huge, power-hungry and low-yield G80 GPU.

    The true high-end replacement family for the 8xxx-series is coming around Q2 of 2008. It has been in design for at least the past year and is NOT just a tweak of the G8x/G9x architecture. If you really HAVE TO upgrade your system right now, just get a SINGLE 8800GT 512. At this point in time, do not invest in SLI. Keep you hands in your pockets and wait for the next gen. A single copy of the high-end version of the next-gen GPU family from nVidia is likely to have more GPU horsepower than dual 8800GTX.
  • Griswold - Thursday, December 13, 2007 - link

    "The true high-end replacement family for the 8xxx-series is coming around Q2 of 2008. It has been in design for at least the past year and is NOT just a tweak of the G8x/G9x architecture."

    Its going to be an evolved (note: thats a fair bit more than just tweaked) G80/G92. You dont design a completely new architecture in a year. Remember what nvidia claimed at launch of the G80? Its been in the works for several years. They will squeeze every bit of revenue out of this architecture before they launch their true next generation architecture (on which at least one team must have been working since the launch of G80).
  • retrospooty - Tuesday, December 11, 2007 - link

    A card with:
    720MHz core clock
    2000MHz shaders
    2400MHz memory
    1GB or memory

    would absolutely rock, so why haven't we got one?

    Ummm.... Wait until the high end card is released in january and then see what the specs are. Its suppsed to be a dual GPU version like the 7950GTX was. So think 2 8000GT SLI performance. The memory wont likely be 2400mhz, but it will be dual channel for 512mbit bandwidth.

Log in

Don't have an account? Sign up now