Originally Posted by raghu78
Even accepting your argument that you want to give 1300 Mhz to GTX 670 the BF3 comparison is wrong.
670 : 58.1 avg, 36 min, normalized avg 61 2560 X 1600 FXAA
7950 : 54.2 avg, 35 min 2560 X 1600 MSAA 2xplease read before you reply
. I am quite sure the 670 will at best tie the HD 7950 and at worst fall slightly behind.
I have no idea what you mean about '1300Mhz to the 670'. I never once mentioned anything even related to that. I said i was going to give a 5% bump to the 670 because it was a reference model being compared to a non-reference with improved cooling and better PCB components. From my experience, and from pretty much every review I've seen (at least with non-faulty non-reference models), a non-reference card of the same make will overclock/perform 5% better. Here
's one such example of how a non-reference card beats a reference card by 7% additional fps. A 5% increase seems perfectly reasonable here.
I defended my position regarding that slight difference with 2xMSAA here
and again here
when the driver difference was brought into question. However, since i made this post today, HardOCP released a new review which tested both games at the exact same settings here
. Lets look at those numbers shall we, as now there's more data available than when i originally wrote this:
7950: 54.7 -- what do you know, it's almost the same as above despite the newer driver and the fact that 2xMSAA is off. Guess my two assumption in the two links above were correct after all?
So you see it makes no difference.
Yes, it says right in the introduction of that review, "AMD's reference Radeon HD 7950 core clock speeds are 800MHz, and XFX took it upon themselves to increase this to 900MHz without the user needing to change anything. "
So, 900/800 = 12.% overclock. This is why i said, i will "increase the performance of the 670 by 12% (5% for being a reference model, 7%
for being 11% underclocked compared to the 7950)." You see i only bumped it up 7% for that 12% overclock because i know 670's have lower scaling then a 7950. That 7% number is actually too low given that the 7950 only scales 10% better than a 670 (so i should have technically used 10.8%) , but i wanted to be generous towards the 7950.
As for the "670 OC (1234) : 67.4 avg, min 33" that data was not available when i wrote this. Look at the time stamps, i wrote this 4 hours before HardOCP put that article up. If that info was available earlier, than i wouldn't have wasted 2 hours writing this comparison. But now that it is available, we can talk about i guess
(but don't try to use it as evidence disproving the original data HardOCP had available to me that i had to extrapolate on). Skip down to after the next quote, i'll talk about this there.
Yes i will agree that a 7970 is about 5%-8% faster than a 7950, i will show you why.
Warning: Spoiler! (Click to show)
If you take the 7950's core clock of 1150 you will need to increase it by 11.3% to get the 7970's core clock of 1280. Since the 7950 only gains .76% fps for every 1% increase in core speed (shown here
), then you take 11.3% and multiply it by .76% which equals 8.5%. So take the 7950's original 74fps and multiply it by 8.5% (1.085) which comes out to 80.3fps. That is what the 7950 would have scored if it was at 1280MHz core. Now take the 7970's fps of 84.9 and divide it by the 7950's new fps of 80.3 and we get a percent difference of 5.7%. Okay but what was the point of all that? I still don't understand your point.
Anyways, now onto the new data you're talking about that i DIDN'T have when i wrote my original post. First thing i notice is how damn close i was of estimating the 670's overclock potential in that game in my original post. I calculated from a stock reference model that it would hit 65.9fps and it ended up being 67.4. That's pretty good if you ask me given the data i had to work with. But that's not relevant now given the new data set, lets look at just that one now:
Oh, i already went into this in this post
. I don't know how far you read in this thread before you posted, but it's usually a good idea to read everything
. I specifically say i don't think the 1050c overclock on the 7950 is fair and that it should be boosted up to 1200c equivalent. This is what i ended up with in that post for quick reference:
7950: 69.6 = ~77 at 1200c
Like i said at the top of my original post. I have no biases, as this clearly proves it. i boosted the 7950 up when it deserved it for the comparison to be fair. I'm in complete agreement that the 7950 kills on Dues EX given this new information, but my original guess/estimate wasn't far off. i said they would be about even in this game when comparing them at low overclocks (7950 at 900mhz vs the 670 at about 940mhz). Given the fact that the 7950 scales 10% better with an overclock (like i showed above), it's not surprising at a higher overclock like this that we find them about 10% apart.
You comfortably forgot to include other games like witcher 2, metro 2033, crysis 2, alan wake in the comparison
I didn't forget, metro 2033 and crysis 2 are down below in my post in the Guru3d section. As for the other games, i would love to compare them but i didn't have the necessary data to compare them fairly. If you have some benchmarks for those games for each card then feel free to post a link and i will look over it.
Where did these numbers come from ?
You forgot to notice the test bed for the XFX HD 7950 is a power hungry 45nm core i7 920 (3.6 Ghz OC) with an extremely power hungry X58 chipset (triple channel memory) based motherboard. The GTX 670 runs on a 32 nm 2600k (4.8 Ghz) with a much less power consuming P67 chipset based motherboard. So direct comparisons will not be accurate.I can accept the GTX 670 is more power efficient but that efficiency comes at the cost of running the risk of being bandwidth constrained
in games like Metro 2033 and more so in the future. I wouldn't even hesitate for a second to buy the HD 7950 given its going to be less affected by bandwidth constraints and well worth the extra power cost.
The numbers came from the power consumption page. I guess you didn't notice they listed the system power consumption too so you could subtract it and find out just the cards power consumption? I didn't forget the 7950 was on a more power hungry system, but that doesn't change anything because i subtracted their consumption. At the top of the graph it says, "Total System Wattage - Without Video Card = 160w." So i took the 529W value and subtracted the 160W value which comes out to 369W for JUST the GPU. I did the same for the 670 and it comes out to 265W. Now take 369w and subtract 265W and we get our power consumption difference of 104W or 28% percent difference (265/369 = .72, now take the inverse of that and convert to a percent = 28%). The power consumption difference comes out to $28.47 per year. it's not huge, but it's certainly relevant and should be part of the decision. This drops the price of the 670 to $372 after the first year when compared to the 7950.
First off, the 7950 is being scaled too high. I said i thought i was scaling it up too much and later on after that new HardOCP review came out today i can easily confirm that now. I was WAY too generous with the numbers here in favor of the 7950. I'm a little disappointed that you pointed out the 7950 was not overclocked enough in the Deus ex example above (and i was in agreement with you even before you posted this) but you don't have a problem with the 670 being too low here? This clearly shows some small degree of bias (or maybe you just overlooked it, i hope it's the later
). Given what i now know, i wouldn't have given the 7950 a 19% boost here, but instead maybe only a 12% boost. So instead of 50fps it should have been around 46fps. So 46fps for the 7950 and 39fps for the 670, or a 18% difference for metro.
Bandwidth has nothing to do with anything here. NVidia has always struggled with Metro because the game just isn't optimized well and NVidia could never get good drivers for it. I have no idea what your gtx 580 example compared to a 6970 has to do with anything. Care to elaborate? Either way, this info has no weight as it doesn't change our measured FPS values. NVidia just sucks at Metro and it always has.
Hard reset is a win for GTX 670. no doubt about that
I don't know what you keep talking about winning or losing. Winning or losing at a certain game is not the point of this exercise. The point is to find the average fps in ALL games so we can use that info tp generalize how good each card is in ANY game now or in the future.
Games like Call of Duty: Modern Warfare 2, FarCry 2 are doing such high fps that the comparison is so inconsequential. How much are you going to able to tell any playability differences at 80 fps compared to 90 fps or 120 fps compared to 130 fps.
There's no such thing as inconsequential numbers in a benchmark. This data is being used, like i just said above, to find the general strength of the card in ANY game by extrapolating all of our data. But just for the sake of answering your question, there is a big difference between 90fps and 120fps for someone with a 120hz monitor (like myself). I can easily tell the difference both visually and by the feel of the game.
Compare that to metro 2033 where there is a 10 fps diff (25%) and Crysis 2 at 2560 X 1600 where there is a 3 fps diff (7.5%). Anno 2070 will run clearly better on HD 7950 OC maxed out compared to GTX 670 OC
Yes, some games run better on AMD than NVidia. It has really nothing to do with the cards themselves but with the drivers. I don't even know why you're so hung up on which games each card does better at. The only time this is of concern is if you ONLY care about performance in a specific game because that's all you play. We're trying to draw a broad conclusion on which card is generally better in all games.
That 670 link is to a weak factory overclocked 670. Again, your confirmation bias is clearly coming to light as you have no problem comparing a light factory-oc'ed 670 to a heavy manually overclocked 7950 but when it was the reverse you had a problem with that (and for the record, so did i, proving that i'm being impartial to the data).So given this data, the 7950 is a reference model and the 670 is a non-reference model, that means i'm going to give the 7950 a free 5% bump (like i did for the 670 earlier when it was in a similar situation).
Guru3D is actually using a non-reference card too, but didn't state that in the article you linked. If you look at the picture of the card with that review you linked
, you can clearly see it's a HIS non-reference mode
BUT we also need to scale the 670 up to the equivalent of a 7950 at 1150c. To do this, refer to my earlier post
(skip to the bottom where i talk about simplify things). So take 1150 and multiply it by 1.09, this gives us 1253, which is the core frequency that the 670 needs to run at to match a 7950 at 1150. This is a VERY easy OC to achieve, as the 670 is averaging well over 1275 on these forums. So we take 1253 and divide by 1137 which gives us the percent difference, which is 10%. Then we take the scaling factor (found here
) which is .69% increase in fps for every 1%, so we take 10% and multiply by .69% and we get
6.9%. So now we need to boost the factory-OC numbers above by 6.9% so we get:
670: 54, 85 ---> 58, 91 at 1253c
7950: 62, 93 at 1150c
So, you can see it's a 7% and 2% difference respectively in favor of the 7950, just like we saw in my earlier extrapolation when i gave the 7950 a 19% additional boost:
Original Anno numbers:
670: 54, 85
7950: 48, 74 = normalized 57, 88
This is a 6% and 4% difference. Almost spot on to the above numbers.
Crysis 2 is a game where Nvidia traditionally excelled. its a TWIMTBP title. the GTX 500 series crushed the HD 6000 series.The HD 7000 series have improved because of major improvements in tesselation performance in HD 7900 series and huge bandwidth which lets them scale better with OCing.
Who cares what it was in the past, we're comparing this generation of cards and right now AMD has an advantage in crysis 2. This doesn't even mater though, as NVidia and AMD trade blows pretty evenly in games based on drivers. The only reason i pointed it out is because Guru3D decided to ONLY include this one game which favors AMD cards which i thought was biased and unfair. They should have included some neutral games or one game that each card excels at (like Batman + Crysis 2, or Skyrim + BF3).
The 7970 is behind here by 5.3% at 1600p and behind at 1200p too by 11%. So i'm not sure how you're drawing the conclusion that the 7970 is better at 1600p. The fact of the mater is, resolution has very little to do with it at 1600p and 1200p. If a game performs better on the 7950 at 1200p then it will almost ALWAYS perform better at 1600p too. The reverse is true for NVidia, if the 670 performs better at 1200p then it will most likely always perform better at 1600p too. This is because it's almost entirely dictated by drivers, if one card is better at a game because of drivers, then it will be better at almost any resolution (especially when you're below 1600p).
You haven't accounted for games where Nvidia gets hammered like Metro 2033, Alan Wake (all demanding games where the performance difference makes playablility difference).
Yes i have, that's why i included every single game in the test set so i can get a good idea of how each card will averagely perform in any game when compared to its counterpart. There's a lot of games NVidia is better at too, actually, it's probably more games than AMD to be honest and that's really the only reason the 670 comes ahead by ~5% when you average everything up. Both cards are about equal in their actual brute strength, they're just being limited by drivers or other game related factors which accounts for this small differential.
hardocp's test suite is limited to 4 games (BF3, Batman AC, Skyrim, Deus Ex) which I feel is not enough to make a really good comparison. Ideally 8 - 10 games released in the last 12 - 18 months is the best way. When we add more games like Witcher 2 , Alan Wake, Crysis 2 it will make things more favourable for HD 7950 OC wrt GTX 670 OC. Anyway thats just my opinion. I look at the most recent and the most demanding games.
Where is your evidence that proves that the 7950 is better in these so called more demanding games? From what i've seen they're pretty even. Feel free to post benchmarks so we can compare though
My opinion is very clear. If you want perf/watt you can go for GTX 670 for all other cases get a good card like Sapphire HD 7950 OC, push your card to the max with voltage OC and enjoy max performance across a wide range of games. Also you will have invested on a more future proof card which won't be bandwidth constrained in games.
The data is pretty clear, both cards are about even (within 2-3% at max overclock at 1600p, slightly in favor of 670), so i don't know how you can conclusively say the 7950 is a better overclocker and performs better across a wide range of games. The fact that that 670 is ahead by 2-3% is evidence that it's slightly better over a wider range of games. There's also no evidence to suggest that the 7970 is more future proof. At 1600p or less the extra 1gb of VRAM will really not have any real-world and measurable impact. Especially considering how we're stuck in the era of weak console ports for the next ~2 years. I also don't know where this 'bandwidth constrained' concept came from, but there's no evidence currently to suggest that the 670 is constrained. There's also no evidence proving that games in the future will be able to make the 670 bandwidth constrained. I welcome you to present the evidence if it does exist, but i haven't seen any
But with that all being said, lets just assume the 670 and 7950 are completely even in terms of average game fps. The 670 is better on every other front. It's audibly quieter by ~96%, it runs cooler (ie, it produced less heat, i'm not talking about its core temps) so it will heat up a room less, it uses less power which saves ~$30 per year., it works with SLI better so that opens up some more viable upgrade routes (though i'm confident AMD will eventually get CF straightened out before it's time to think about a second card), and it has slightly less micro-stuttering and dynamic Vsync. But when you consider all that and then realize that that 670 is a hair faster too, it really becomes a no-brainer.
Enjoy the Great Wall bro
Edit: Now that Brett has posted his info
, we know the scaling for the 670 is 81% not 69% like i used to calculate various things in this post. So that means most of my extrapolated data for the 670 is about 12% less than it should be. So keep that in mind before i get a chance to fix this post sometime tomorrow.Edited by SeanPoe - 5/15/12 at 2:51am