Price/Performance Analysis

We decided not to use the MSRP (manufacturer suggested retail price) in our value calculation because people are buying the cards at market prices, not suggested prices. To answer the question "what card should I buy right now?" we have to evaluate the situation as it stands.

As a reminder, here is the street pricing data that we gathered and estimated (all prices are USD).

NVIDIA GeForce FX 5950: $380
NVIDIA GeForce 6800: $300
NVIDIA GeForce 6800 GT: $410
NVIDIA GeForce 6800 Ultra: $540
NVIDIA GeForce 6800 Ultra Extreme: $600

ATI Radeon 9700 Pro: $180
ATI Radeon 9800 XT: $400
ATI Radeon X800 Pro: $420
ATI Radeon X800 XT: $540
ATI Radeon X800 XT Platinum Edition: $600

When applying a graph to some set of data, it becomes more tangible and concrete, almost more "real" in a sense. But prices always fluctuate in the market, drivers come out that change performance, and people attach differing values to certain aspects of a product. The value graphs we have included are only one indicator in a sea of data. But making decisions between cards is difficult in very close race we have, and any indicator we have helps.

Interestingly, the 9700 Pro puts in a strong showing. Of course, in determining what purchase to make, we need to take into account minimum acceptable performance levels as well. The 9700 Pro may be a good value for many games, but it just won't deliver the frame rates in current and future titles, at the resolutions to which people are going to want to push their systems.

If something on the performance level of a 9800 XT or 5950 Ultra is desired, the NVIDIA GeForce 6800 is the way to go, every time. It's cheaper than both previous generation top-of-the-line cards, and offers better consistent performance and scalability than both. Of course, we haven't taken a look at a comparable product from ATI yet, but if it performs on par with the 6800, we'll be in good shape here. We are hearing rumors of an X800 SE part, but we have no confirmation from ATI on this. We will definitely re-evaluate the situation when that card comes to market, but for now we are very impressed with the performance and value of the stock 6800 part.

For higher performance, which is a better value? Should you go with the 6800 GT or X800 Pro? Is it a choice between the 6800 Ultra and X800 XT? Well, the answer is: go with the card that gives you the best value for the games that you like.

If that's too difficult a decision to make, our data shows the 6800 GT to lead the X800 Pro in performance. Maybe the $10 is a factor to you and maybe it isn't, but in this close race, price should definitely be a determining factor.

We can also see the 6800UE leading the X800XTPE most of the time as well. Gamers who choose these very high performing (and priced) cards will likely place less value on "value" than on how well the card does under on particular game. In those cases, check the performance graph for the game of choice and go for it.
X2: The Threat Performance Final Words
Comments Locked

46 Comments

View All Comments

  • TrogdorJW - Tuesday, July 13, 2004 - link

    Regarding the logarithmic scale, again, I don't think that it was a really bad idea. It tries to equalize scores that really have a lot of variables. What would be the best way of measuring bang-for-the-buck? Well, here's what I think would actually have to come into consideration.

    Well, you would have to take into account the desired average frame rate. This of course varies from game to game, but a score of somewhere between 60 and 100 FPS would probably be ideal. Of course, this would vary between games, as a game like Unreal Tournament 2004 "needs" higher frame rates more than something like Flight Simulator 2004 (or whatever we're on now).

    Why 100, when you can't really "see" things that fast? Well, ideally, you would want V-synch enabled for maximum image quality. 100 FPS without V-synch would probably get pretty close to your refresh rate, i.e. 85 Hz, with V-synch. Meanwhile, 60 FPS with an 85 Hz refresh rate might end up scoring more like 42.5 FPS with V-synch enabled. Even worse, 59 FPS with a 60 Hz refresh rate might score as low as 30 FPS.

    Anyway, the scale would be weighted, so scores lower than this threshold would be punished increasingly more, while scores above this threshold give diminishing returns. In other words, if we're shooting for 60 FPS, a score of 55 is pretty close, so it gets maybe 89% on a relative scale (instead of the mathematical 91.7%), while a score of 50 might only get 75% (instead of 83.3%) and a score of 45 might get 55% (instead of 75%). Meanwhile, 70 FPS might score 110% (instead of 116.7%) and 80 might score 115% (instead of 133.3%). That's just a rough example to illustrate what I'm talking about - no actual formula was used there.

    You might want to consider *minimum* frame rates (although the weight given to the average frame rate does accomplish *some* of this). So maybe we want a minimum frame rate of 40 FPS, and again its weighted so that dropping below this hurts more, while exceeding this value benefits less. And maybe there should be some factor taking into account the percentage of time that a card drops below this threshold?

    Price is also a factor. Some people have limits, so you would have an inversely weighted price scale based off your target. Say you want to spend $200. Cards costing less than that would be more desirable, even with lower performance, while cards costing more would be less desirable.

    You can go on with all sorts of ideas. In the end, though, it ends up being far too complex for anyone but a mathematics professor. So, the log scale rating that was used isn't terrible... it's just a little interesting, since a straight scale is more easily calculated, more people are familiar with it, and some would say it’s more “accurate” (although that last is more opinion than anything).
  • jiulemoigt - Monday, July 12, 2004 - link

    hummm also Doom3 is confirmed for quakecon so the next game to kick your pc's (whatever) is due to come out next month should be interessting to see if any of the value cards can handle it.
  • DerekWilson - Sunday, July 11, 2004 - link

    our grphing engine doesn't display log scale. I went to all this trouble to get around that :-)

    Also, we will get color coding in our graphing engine sometime soon, but we don't have it yet ...

    Normalizing everything to avg fps and avg cost wouldn't be as useful if done on a per graph basis (which it would have to be done on)... Unless you picked an arbitrary normalization point like desireable fps across all games and desireable cost for that fps .... but that's too sticky and specific ... You have a good idea, but i don't think it would be straight forward to impliment --- we'll see what we can do with this though, thanks.
  • Pumpkinierre - Sunday, July 11, 2004 - link

    Normalise all the results to the average fps/cost of all the cards. Any card close to the normal average(1 or 100 if x100) will be a reasonable to good buy and the outer extremes will be accentuated while still being contained (unlike logarithmic representation that distort the comparison).
  • trexpesto - Saturday, July 10, 2004 - link

    Hmm, can your graph generator show a log scale? Then maybe you can use the raw fps/$ numbers but it would be at the display level that the scaling was done.
  • trexpesto - Saturday, July 10, 2004 - link

    Not sure I like the log scaling.
    Looking at http://anandtech.com/video/showdoc.aspx?i=2113&...
    in the first graph with no AA/AF the scores for 9800XT vs 6800NP are 12.2 and 14.9 respectively.

    So:
    9800XT 66fps/$400 = .165 AT-score = 12.2
    6800NP 93fps/$300 = .31 AT-score = 14.9

    So the fps/$ is almost double but the AT-score is about one-fifth increase.

    Not that anyone will ever pay 400 for an 9800XT again - vendors are you listening?
  • AtaStrumf - Saturday, July 10, 2004 - link

    You might wonna improve the presentation of your results (at least in the Final Words page) by putting them in as few as possible tables (and add/or only use relative results (%) beside actual FPS (mandatory for the GPU which represents 100% or 1,00), to better relate the difference in perfomance -- now being represented by graph bars length which is not that accurate) instead of hundrets of graphs on tens of pages. It makes getting a good overview and doing quick comparisons very hard. You can also use some creative color coding (green = good, red = bad, etc.) to futher increase the readability of these tables. You have done a bit of this in some of your more recent reviews and I'd like see you do it even more and even step it up a notch or two.
  • DerekWilson - Saturday, July 10, 2004 - link

    as for the log scale ...

    if we had said:

    Value = fps / price

    we would have had tiny values which look really bad on our graphs :-) So our first option was to scale the graphs up by multiplying by 100:

    Value = 100 * fps / price

    The problem with this is that the visual impact of the scaling seemed too dramatic. Now, we don't mean with the high priced ultrasuperextremegold cards. The problem was the difference between the X800 Pro and the 6800GT ... The GT certainly comes out on top, but we are talking about current market prices and a difference of $10. We didn't want to give this difference (the cost of a couple value meals at the local fast food place) more weight than people might ascribe it.

    Of course, there are infinite other ways to look at this. Is frame rate *REALLY* important to you? Then its perfectly valid to look at this graph:

    Value = fps^2 / price

    or is cost the absolute driving factor to you?

    Value = fps / price^2

    Its all relative really, and very hard to determine. We just wanted to show which cards will give you more frames per dollar, but we didn't want to ascribe a dollar value to every frame ... So the log scale seemed to be our best bet.

    But we are definitely open to feedback on how we approach value from a quantitative standpoint. Please let us know what you think, or if there is a better way to approach the issue.

    Thanks!
  • DerekWilson - Saturday, July 10, 2004 - link

    Warcraft III is an interesting game. It is a DX8 game, but OpenGL rendering can be enabled with a commmandline switch. In doing some testing, we noticed that OGL mode gets MUCH higher frame rates on both cards (if you talk to some hard core wc3 freaks out there, they might already enable OGL mode to keep framerate up).

    We spoke with Blizzard about the differences between the DX8 and OGL rendering paths. Specifically, we wanted to know if there was a quality difference that would make the OGL path run faster than the DX8 path. Blizzard informed us that the two paths had converged on the same quality level. They would not say that the DX8 and OGL paths rendered the exact /same/ image, but they maintained that the only reason they didn't was that DX8 and OGL can't rendere the exact same pixels in certain situations.

    We may taclke some OGL benches in WC3 in the future if people are interested. But since we're trying to show realworld performance in these benches, and DX8 is what most people will use, we haven't felt like including these numbers would be useful for now.

    Blizzard didn't say anything about capping framerates in DX8, but, then, we didn't ask that question specifically.
  • Marsumane - Friday, July 9, 2004 - link

    Actually WCIII is DX8 based. Also, wc3 is quite odd with graphic cards. Even with a high end cpu, on toms hardware, the cards dont exactly scale evenly as they do in other benchmarks. Its like theres a cap placed on them in the engine or ATI's drivers dont have some sort of cap on them. I believe it is in the engine due to all cards seeming to have similar frames dispite the card used (unless AA or AF is turned on). The ati cards have a slightly lower cap it seems. I mean the difference is like 3fps between the 9700p and the 6800U. It makese no sense why they all stop here.

Log in

Don't have an account? Sign up now