GPU Showdown: Fermi Vs. Cypress - Both Overclocked on Air

[blanketyblank]

I think the only important tests are those where one of the cards was getting under 60 fps at least at this resolution. For example in Left 4 Dead 2, I don't think there is any noticable difference between 268 fps of the 5970 vs the 140 fps of the GTX 465.

That being said I wish he would have tested different resolutions i.e. with a 30" monitor or dual monitors or something for the games where the fps were above 60.

 

[Daedalus685]

I am wondering how much juice those Fermis require to OC that high, not to mention the heat coming off of them. Impressive, though.

Also, it really shows how much the ATI cards are in need of a price cut, and what I've said for months and months now, if you're will to OC there's no reason at all to buy a 5870 over a 5850.

Note that in the tests the fellow was unable to clock his 5850 that high. He used the same 5870 with a different bios to disable shader cores to 5850 levels.

The OC value of the 5850 is representative of a 5870, not a 5850.. It is merely the results one would get 'IF' the 5850 could clock that high.

Same goes for the 470 which was an unlocked 465.

That being said.. the extra cores in the 5870 seem to accomplish almost nothing for whatever reason.. Which has been shown time and time again.. Hopefully this can be fixed for the 6000s or the high end may not be much faster at all.

 

[Avalon]

Very cool benches.

 

[Grooveriding]

Awesome set of benches! Credit to him for all that thorough work.

Having used 5870 crossfire and 480 SLI, having used the 5870 crossfire setup on a 1920x1200 screen and a 2560x1600 and the 480 SLI solely on the 2560x1600, these are the kind of benches I like to see.

I don't think it's being discussed enough, that with the amount of GPU horsepower available to us today, and the demands placed on GPUs from games, we've hit a plateau where we have more performance than we need.

The only exception I see to this are 30" and eyefinity & nvidia's version of eyefinity resolutions. I think that ATI & Nvidia are likely seeing this as well and things like eyefinity, 3dvision, physx etc, are their way of trying to keep buyers justifying the need to upgrade.

As it stands currently, there is nothing you cannot max out at 1920x1200 with a 5870CF or 480SLI setup. And close to nothing that you cannot max out at 2560x1600 with a 480SLI setup, short of turning on the depth of field effect in Metro 2033.

Game developers need to step it up or we'll keep hearing about near useless features like physx and 3dvision etc.

Excellent benches for people who plan to overclock.

 

[KIAman]

Note that in the tests the fellow was unable to clock his 5850 that high. He used the same 5870 with a different bios to disable shader cores to 5850 levels.

The OC value of the 5850 is representative of a 5870, not a 5850.. It is merely the results one would get 'IF' the 5850 could clock that high.

Same goes for the 470 which was an unlocked 465.

That being said.. the extra cores in the 5870 seem to accomplish almost nothing for whatever reason.. Which has been shown time and time again.. Hopefully this can be fixed for the 6000s or the high end may not be much faster at all.

I think something went wrong on that because with 10% less shaders, the 5850 was only within 1-5% of the HD5870 on most benchmarks.

 

[Insomniator]

Awesome set of benches! Credit to him for all that thorough work.

Having used 5870 crossfire and 480 SLI, having used the 5870 crossfire setup on a 1920x1200 screen and a 2560x1600 and the 480 SLI solely on the 2560x1600, these are the kind of benches I like to see.

I don't think it's being discussed enough, that with the amount of GPU horsepower available to us today, and the demands placed on GPUs from games, we've hit a plateau where we have more performance than we need.

The only exception I see to this are 30" and eyefinity & nvidia's version of eyefinity resolutions. I think that ATI & Nvidia are likely seeing this as well and things like eyefinity, 3dvision, physx etc, are their way of trying to keep buyers justifying the need to upgrade.

As it stands currently, there is nothing you cannot max out at 1920x1200 with a 5870CF or 480SLI setup. And close to nothing that you cannot max out at 2560x1600 with a 480SLI setup, short of turning on the depth of field effect in Metro 2033.

Game developers need to step it up or we'll keep hearing about near useless features like physx and 3dvision etc.

Excellent benches for people who plan to overclock.

The fact that you can max out everything with $800-$1000 worth of video cards is not a sign that developers need to step up.

90% of people are playing with a 5770 or less... and while i agree graphics have looked pretty stagnant for a while I'd just like to see midrange cards play freakin crysis better already.

 

[Daedalus685]

I think something went wrong on that because with 10% less shaders, the 5850 was only within 1-5% of the HD5870 on most benchmarks.

This kind of poor high frequency behavior has been shown elsewhere many times before. I don't why it happens but it seems to be that the current 5000 series shaders do not scale well at high clocks.

None have been this highly over clocked but a 1ghz 5870 is not much faster than a 1ghz 5850.

 

[Grooveriding]

The fact that you can max out everything with $800-$1000 worth of video cards is not a sign that developers need to step up.

90% of people are playing with a 5770 or less... and while i agree graphics have looked pretty stagnant for a while I'd just like to see midrange cards play freakin crysis better already.

I hear that and agree on the point that with advancements in graphics being stagnant midrange cards should be doing better at this point.

In the past until this generation, even with expensive graphics card setups, being able to run the latest games on their highest settings with AA was still not a possibility, but since the release of the 58XX and 480/470 series this has become a reality.

You can get pretty hefty performance from multi card setups of the midrange contenders.

 

[Bill Brasky]

I realize you can't really compare numbers between different benchmarks, but I still think this is interesting. NCspecV81 benched at 1920x1080, while Anand benched at 1920x1200. Even though NCspecV81 used a faster cpu and OC'ed the gpu's, his HD5970 was on ave. only 1ps faster than Anand's in BC2. Have I missed something here? Perhaps different quality settings I didn't see?

edit: Although NCspecV81 didn't specify which scene he used and had 8xAA, so nevermind.

 

[SolMiester]

Wow, OC 480 = 5970?!

Imagine if Fermi had come out at original cores and clocks?!, it would indeed have been equal to the ATi sandwich!

 

[ModestGamer]

Note that in the tests the fellow was unable to clock his 5850 that high. He used the same 5870 with a different bios to disable shader cores to 5850 levels.

The OC value of the 5850 is representative of a 5870, not a 5850.. It is merely the results one would get 'IF' the 5850 could clock that high.

Same goes for the 470 which was an unlocked 465.

That being said.. the extra cores in the 5870 seem to accomplish almost nothing for whatever reason.. Which has been shown time and time again.. Hopefully this can be fixed for the 6000s or the high end may not be much faster at all.

could just be software not efficienctly using hardware. Not like its never happened before.

 

[3DVagabond]

I remember when Cypress was first released sites found you could O/C the RAM higher than it was truly stable at because it has some type of error correction that will stop it from having artifacts or giving any visual clues, but adversely affect performance. Is it possible the RAM is actually O/C too far and adversely affecting performance?

 

[tviceman]

I realize you can't really compare numbers between different benchmarks, but I still think this is interesting. NCspecV81 benched at 1920x1080, while Anand benched at 1920x1200. Even though NCspecV81 used a faster cpu and OC'ed the gpu's, his HD5970 was on ave. only 1ps faster than Anand's in BC2. Have I missed something here? Perhaps different quality settings I didn't see?

Yes there was a quality setting you missed. Anand didn't use 8X AA, whereas this particular benchmark in BFBC2 did.

http://www.overclock.net/overclock-...pu-showdown-fermi-vs-cypress.html#post9891837

 

[MrK6]

Wow, OC 480 = 5970?!

Imagine if Fermi had come out at original cores and clocks?!, it would indeed have been equal to the ATi sandwich!

Well, FWIW, that tester only retested games where the GTX 480 was already very close to the 5970. The two cards are so different that there performance differences are entirely application based, and really are all over the place.

 

[Keysplayr]

Well, FWIW, that tester only retested games where the GTX 480 was already very close to the 5970. The two cards are so different that there performance differences are entirely application based, and really are all over the place.

So you're saying that all the games that were not tested would show the AMD cards in a much better light. We'll I guess that's pretty much the only thing that could have been said in this case. Fermi scales a lot better with overclocking than the 5xxx series. And I believe that was shown in a much broader spectrum of games than shown here. In many games that had shown stronger on 5xxx cards over Fermi, it sometimes turned that table when compared cards were o/c'd. 5xxx series cards seem to require a much greater percentage in an overclock to achieve the same gain as Fermi in a given title.

 

[Grooveriding]

So you're saying that all the games that were not tested would show the AMD cards in a much better light. We'll I guess that's pretty much the only thing that could have been said in this case. Fermi scales a lot better with overclocking than the 5xxx series. And I believe that was shown in a much broader spectrum of games than shown here. In many games that had shown stronger on 5xxx cards over Fermi, it sometimes turned that table when compared cards were o/c'd. 5xxx series cards seem to require a much greater percentage in an overclock to achieve the same gain as Fermi in a given title.

Fermi does scale better, but about 1% of 480 owners are going to get an overclock of 980 on the core like is shown in that review. 800 is the norm, unless you want to ramp voltage, then 850 or so, of course with added voltage Fermi turns into an absolute furnace as it's already hot enough as is at stock.

 

[Keysplayr]

Don't worry about the heat. The gpus can obviously take it. My point, which I think you missed here, is that Mhz to Mhz, Fermi scales better than 5xxx. Nothing to do with a GTX480, or 1% of GTX 480 owners reaching 980Mhz. Just wanted to clear that up.

 

[Daedalus685]

Don't worry about the heat. The gpus can obviously take it. My point, which I think you missed here, is that Mhz to Mhz, Fermi scales better than 5xxx. Nothing to do with a GTX480, or 1% of GTX 480 owners reaching 980Mhz. Just wanted to clear that up.

The scaling of cypress isn't linear at all though. IT scales worse (far worse) at higher clocks but seems to be about the same up to around stock. Whatever they changed in the shaders to add DX 11 kills performance at higher clocks.. I'd wonder if the error correction is easier to get to than people think. It is just a dead dog at over 900mhz. Though the error correction doesn't explain why the extra shaders in a 5870 accomplish perhaps 5% more performance over a 5850.

 

[ArchAngel777]

I suspect the refresh of Fermi will be something based around the GF104. I was really surprised at how they managed to tackle the performance\watt with a significant degree with the GF104.

I agree. That was some great engineering on their part. I was extremely skeptical over GF104, but they really delivered. Now if we can get a GF104 type 480 with 512 SPs and clock her way up there... ATM, the GTX 460 is an absolutely awesome card for the money, BUT... For those of us with a 280/285 already, it doesn't really compell us for an upgrade.

 

[Genx87]

The scaling of cypress isn't linear at all though. IT scales worse (far worse) at higher clocks but seems to be about the same up to around stock. Whatever they changed in the shaders to add DX 11 kills performance at higher clocks.. I'd wonder if the error correction is easier to get to than people think. It is just a dead dog at over 900mhz. Though the error correction doesn't explain why the extra shaders in a 5870 accomplish perhaps 5% more performance over a 5850.

Doesnt Fermi have more cache available to the shaders? Also isnt Fermi MIMD where Cypress is SIMD? Could be at those speeds Fermi simply has the capability to feep its units fed with instructions at a much higher efficiency. I guess one way to test this is to see if GF104 scales in a similar way. They scaled back the cache on the GF104

 

[Daedalus685]

Doesnt Fermi have more cache available to the shaders? Also isnt Fermi MIMD where Cypress is SIMD? Could be at those speeds Fermi simply has the capability to feep its units fed with instructions at a much higher efficiency. I guess one way to test this is to see if GF104 scales in a similar way. They scaled back the cache on the GF104

It could be, but I'd have thought that cache and MIMD would affect program scaling (i.e. fermi woudl be crazy fast in some applications) but not frequency scaling. I think it is more that cypress is bad at scaling more than fermi is good at it. Not to belittle how exceptional teh scalign of Fermi is.. just that the 5870 is very poor.

 

[MrK6]

So you're saying that all the games that were not tested would show the AMD cards in a much better light. We'll I guess that's pretty much the only thing that could have been said in this case. Fermi scales a lot better with overclocking than the 5xxx series. And I believe that was shown in a much broader spectrum of games than shown here. In many games that had shown stronger on 5xxx cards over Fermi, it sometimes turned that table when compared cards were o/c'd. 5xxx series cards seem to require a much greater percentage in an overclock to achieve the same gain as Fermi in a given title.

Well, I think that was the point of his tests, to say "check it out, a single GPU can surpass a multi-GPU solution in these titles." You'll notice he didn't show a comparison in AVP, Crysis or CS:S, where there's no way Fermi would have caught the 5970. The Fermi architecture scales better than Cypress in newer titles because it's a newer architecture. The fact that AMD is still competing with a 3+ year old design is impressive in itself.

Don't worry about the heat. The gpus can obviously take it. My point, which I think you missed here, is that Mhz to Mhz, Fermi scales better than 5xxx. Nothing to do with a GTX480, or 1% of GTX 480 owners reaching 980Mhz. Just wanted to clear that up.

I don't think anybody is debating that point though. However I would worry about the heat. I'll add to Grooveriding's point that I also think it's much more realistic to run a 5850/5870 at 1GHz rather than a GTX 480 at 850MHz when environmental variables like heat and noise are considered.

The scaling of cypress isn't linear at all though. IT scales worse (far worse) at higher clocks but seems to be about the same up to around stock. Whatever they changed in the shaders to add DX 11 kills performance at higher clocks.. I'd wonder if the error correction is easier to get to than people think. It is just a dead dog at over 900mhz. Though the error correction doesn't explain why the extra shaders in a 5870 accomplish perhaps 5% more performance over a 5850.

There's always a possibility that the overclock isn't stable (it is pretty extreme for air), but I think it's also likely that there's some architecture bottleneck that overclocking simply can't overcome and is holding back performance. Maybe someone with more knowledge of GPU architectures can postulate what's going on here.

 

[Dribble]

I agree. That was some great engineering on their part. I was extremely skeptical over GF104, but they really delivered. Now if we can get a GF104 type 480 with 512 SPs and clock her way up there... ATM, the GTX 460 is an absolutely awesome card for the money, BUT... For those of us with a 280/285 already, it doesn't really compell us for an upgrade.

+1

I think that is the part many of us are waiting for. As a GTX 260 owner who wants a properly large upgrade (last was 8800GT -> high stock o/c GTX 260) not sure GTX 460, or even a 5850 are quite worth it. A GTX 470 is a bit expensive, hot and power hungry.
Now a GTX 475 with using a 512SP GF104 (448 enabled) would be faster, lower power, cheaper (chip would be smaller then GF100), and over clock better. That might just tempt me to spend.

You can tell you lord's and master's that Keys

 

[Grooveriding]

Don't worry about the heat. The gpus can obviously take it. My point, which I think you missed here, is that Mhz to Mhz, Fermi scales better than 5xxx. Nothing to do with a GTX480, or 1% of GTX 480 owners reaching 980Mhz. Just wanted to clear that up.

I got your point, my post began with 'fermi does scale better'

My point is that, to draw attention to a 480's performance with a 1ghz core is only appropriate in the context of it being for show and not a realistic example of what you can expect to get out of one.

The 480s get really hot, and when I see examples of people ramping voltage to get crazy overclocks on them, they are using good watercooling, an additional $500 investment, or if on air, running the fan at 100%, which is loud enough to hear on a different floor of your house.

Just saying that yes the performance is amazing at that speed, but don't expect to actually get those kind of results if you go out and pick one up. Most of them cannot even get to those clocks, extra voltage or not. It's basically saying look at what -could- be possible, if these cards could deliver this speed in a fashion that was realistic as the norm, rather than the exception.

 

[3DVagabond]

It's basically saying look at what -could- be possible, if these cards could deliver this speed in a fashion that was realistic as the norm, rather than the exception.

It's pure porn though to nVidia fans. They can't help but get all giggly about it.