HD 7970: Bulldozer vs. Sandy Bridge vs. Nehalem

[W1zzard]

You know what W1zzard probably has no idea what he is talking about I mean he only runs a forum with 65 thousand members and 2.4 million posts, not to mention some of the most in depth reviews on the internet.

while i appreciate your comment, i disagree. i dont run a religious outfit, so feel free to ask questions and criticize. we're all here to learn

 

[cdawall]

while i appreciate your comment, i disagree. i dont run a religious outfit, so feel free to ask questions and criticize. we're all here to learn

Very true. Just out of curiosity why do you still test at such low resolutions? Just to get a comparison with some of the lower end cards or is there a deeper reasoning?

 

[W1zzard]

Very true. Just out of curiosity why do you still test at such low resolutions? Just to get a comparison with some of the lower end cards or is there a deeper reasoning?

comparison data for low end cards.

in theory i could leave out the graphs of the lower resolutions for high-end cards. i'd still have to bench them for comparison. but some readers might be interested in the low-res graphs to look at them for advanced concepts like cpu dependency, resolution scaling etc.

the majority of readers should have no issues skipping over a few graphs on each page

 

[Tatty_Two]

i am still surprised how well the nehalem processors are holding up especially because of how old they are. I upgraded from my 920 at the start of last year to a 2600k only because it seemed more fun and it ran much cooler.

Ditto^^^ it makes me feel kind of good even though I have had this chip for 3 odd years.....and as most of us here overclock, I would guess that if all the CPU's on test were cranked up to BD's stock clocks, the results would have been even more interesting.

 

[Makaveli]

comparison data for low end cards.

in theory i could leave out the graphs of the lower resolutions for high-end cards. i'd still have to bench them for comparison. but some readers might be interested in the low-res graphs to look at them for advanced concepts like cpu dependency, resolution scaling etc.

the majority of readers should have no issues skipping over a few graphs on each page

Hey Wiz have you seen techreports new approach to graphing FPS if so what do you think about it?

 

[antuk15]

doesn't that mean "tricky" ? actually your statement confirms that.

how do you propose to reliably measure minimum fps? how to ensure decent accuracy? what resolution and accuracy for the measurement do you consider acceptable?

how is minimum fps defined? (just one frame? over one second?)
when does a frame start and end anyway? what about the time between frames?

everybody who shows minimum fps in their reviews uses fraps.

Crysis
Crysis 2
Stalker series
Far Cry 2
Cryostsis
Batman : AA
Metro 2033

There's HUNDREDS of games that have built in benchmark programs that messure minimum fps without the need for fraps or any other program.

A few other website actually provide minimum frame rates but if you look closely its only for games that have built in benchmarks that show and record them. These built in benchmarks are much more consistent then fraps and you should know that.

But yes you're right, They all use fraps

 

[OOZMAN]

Crysis
Crysis 2
Stalker series
Far Cry 2
Cryostsis
Batman : AA
Metro 2033

There's HUNDREDS of games that have built in benchmark programs that messure minimum fps without the need for fraps or any other program.

A few other website actually provide minimum frame rates but if you look closely its only for games that have built in benchmarks that show and record them. These built in benchmarks are much more consistent then fraps and you should know that.

But yes you're right, They all use fraps

Hey man, fraps is awesome.

 

[W1zzard]

Crysis
Crysis 2
Stalker series
Far Cry 2
Cryostsis
Batman : AA
Metro 2033

you want me to bench with those games only? almost all new titles dont have benchmarking functionality

These built in benchmarks are much more consistent then fraps and you should know that.

how do you define consistent ? and whats your reference value to compare to ? part of the issue is what i mentioned further above for which you apparently have no answers. another problem is that time measurements are quite difficult to do on windows. i'd expect fraps to do better in that department than most engine benchmarking code

Hey Wiz have you seen techreports new approach to graphing FPS if so what do you think about it?

just looked at it. good to see someone trying new things.
http://techreport.com/articles.x/22192/11

personally i think fps graphs are too complicated for many readers and offer little additional insight.
not sure why tr graphs their data the way they do, but frame number on the x axis seems like a bad choice. you want to put time on x axis. look how each of their graphs has a different number of frames for its own run.
frametimes on y is also counterintuitive to what most readers expect, especially if the values are in the 20-100 range where people instantly think fps

the use of 99th percentile frametime makes no sense to me (yes i know what 99th percentile is). most people will look at that graph with the big scientific name, skip it, and be impressed with it

time spent beyond 50 ms: good idea. bad naming, i thought FPS again
so each of their benchmark runs runs a different time duration. then they add up how long the frametimes were 50+ ms (for a different number of frames in each run) and then compare these values by putting them in a graph. so they compare a shorter maximum time with a longer time?

edit: so i found their article explaining the changes: http://techreport.com/articles.x/21516
good read. their choices make more sense now. need to think more about it, but that alone is a problem. review readers dont want to read an instruction manual for the review

 

[someone]

Hi there

@techreport
I generally like their ideas very much. As a former user of a multi-gpu setup I particularly know that fps don't tell the whole story. In fact, they don't even tell half the story sometimes. I remember playing Crysis at 70-80fps (vsync=off) which stuttered big time and felt worse than 40fps without micro stuttering.

As per the metrics, they should definiitely normalize everything, that would eliminate the problem of comparing runs of different lengths (or different amounts of frame times). Instead of counting the number of frame times larger than a given threshold, they would then report the proportion which is larger than 50ms. I find this transformation quite standard and straightforward, and it's a clear improvement IMO.



@BD vs. SB vs. NH
The results are pretty much as expected (sadly for AMD, one should note).

Here are two interesting things in the results:

1. In Skyrim, going from 1024 to 1280 keeps the fps constant, the same for 1680 to 1920.
But there's a difference between the upper resolutions and the lower ones:

(black arrow=no difference, red arrow=difference)


2. In Starcraft, on the other hand, there are neither "horizontal" nor "vertical" changes, or at least the upper resolutions are more similar to the lower ones than in our example before (Skyrim):

I suspect the aspect ratio makes the difference, since the upper resolutions are 4:3 and 5:4 whilst the lower ones are 16:10 and 16:9 respectively. The wide screen aspects require more rendering in the horizontal than in the vertical, compared to the 4:3 (and 5:4) ratio.


The explanation, why there is a difference in Skyrim and (almost) none in Starcraft is then, that
- there's not so much going on in the vertical in Skyrim: floor texture and sky texture, from which especially the latter is very simple to compute for the CPU. So, when you add more horizontal pixels, I would expect a much larger amount of CPU computations necessary than when adding vertical ones (just paint some more heaven and floor, to oversimplify).

- in Starcraft, since it's a top-down view, adding horizontal pixels and adding vertical ones should make (almost) no difference. It's far less asymmetric than Skyrim, since more terrain, buildings and units will be computed, regardless in what direction the image is expanded.

 

[HTC]

you want me to bench with those games only? almost all new titles dont have benchmarking functionality



how do you define consistent ? and whats your reference value to compare to ? part of the issue is what i mentioned further above for which you apparently have no answers. another problem is that time measurements are quite difficult to do on windows. i'd expect fraps to do better in that department than most engine benchmarking code


just looked at it. good to see someone trying new things.
http://techreport.com/articles.x/22192/11

http://img.techpowerup.org/120103/Capture283.jpg
personally i think fps graphs are too complicated for many readers and offer little additional insight.
not sure why tr graphs their data the way they do, but frame number on the x axis seems like a bad choice. you want to put time on x axis. look how each of their graphs has a different number of frames for its own run.
frametimes on y is also counterintuitive to what most readers expect, especially if the values are in the 20-100 range where people instantly think fps

the use of 99th percentile frametime makes no sense to me (yes i know what 99th percentile is). most people will look at that graph with the big scientific name, skip it, and be impressed with it

time spent beyond 50 ms: good idea. bad naming, i thought FPS again
so each of their benchmark runs runs a different time duration. then they add up how long the frametimes were 50+ ms (for a different number of frames in each run) and then compare these values by putting them in a graph. so they compare a shorter maximum time with a longer time?

edit: so i found their article explaining the changes: http://techreport.com/articles.x/21516
good read. their choices make more sense now. need to think more about it, but that alone is a problem. review readers dont want to read an instruction manual for the review

Interesting read: hadn't seen that approach before.

If i may make a suggestion: instead of using seconds, why not use the next value along the line, as in tenths of seconds? Wouldn't this catch more of the issues described in that techreport site's article?

Ofc, i'm assuming there's a tool that can measure this because, otherwise, no point in even trying.

 

[MilesRdz]

A lot of people seem to not notice or care that Bulldozer is vastly underutilized in most of these games.
While SB is using half or more of it's resources, BD is using about 1/4th.

If games used more threads to feed the graphics card, we wouldn't be discussing this issue to death.

Some people could argue that if BD had better single-threaded performance, games would run better.
That is true, but it doesn't change the fact that BD is underutilized.

 

[Mussels]

A lot of people seem to not notice or care that Bulldozer is vastly underutilized in most of these games.
While SB is using half or more of it's resources, BD is using about 1/4th.

If games used more threads to feed the graphics card, we wouldn't be discussing this issue to death.

Some people could argue that if BD had better single-threaded performance, games would run better.
That is true, but it doesn't change the fact that BD is underutilized.

and thats why people still have hope for the two-patch solution MS is working on for windows 7 and BD.

 

[ensabrenoir]

and thats why people still have hope for the two-patch solution MS is working on for windows 7 and BD.

Yeeeahhh but ain't

1/4 of 8 =2

1/2 of 4 =2

Sooo.....its pretty much a fair match up ratio

Wait.....no it aint....sorry my math was off

 

[RejZoR]

Good to see my "outdated" Core i7 920 is not for old junk just yet...

 

[TheGuruStud]

and thats why people still have hope for the two-patch solution MS is working on for windows 7 and BD.

That doesn't fix apps only using a couple threads, though.

And we all know that since M$ released the patch it doesn't do anything at all. We will have to wait for win8 for any small improvement.

A lot of people seem to not notice or care that Bulldozer is vastly underutilized in most of these games.
While SB is using half or more of it's resources, BD is using about 1/4th.

If games used more threads to feed the graphics card, we wouldn't be discussing this issue to death.

It wouldn't be discussed at all b/c intel would lose in every game. That's how the game works don't you know?

A lot of sites (paid by you know who) won't bench apps favorable to AMD b/c they're paid not to. We end up with reviews massively one siding a situation regardless of real performance.
I remember Athlon 64s losing to pentium 4s back in the day in benches....now, I wonder how that was possible LOL

 

[BlackOmega]

Where do people get off saying BD is fail? I looked through every page and, more often than not, it was beating Nehalem, and sometimes even beating the 2500k.

While sure it's not superdooperawesome like a lot of people were hoping for, but it's far from fail IMO.

EDIT: W1zzard, how come you guys don't test in 1080p? I'd dare say that the probably THE MOST common resolution used these days, yet it's always omitted in tests.

 

[Aquinus]

Where do people get off saying BD is fail? I looked through every page and, more often than not, it was beating Nehalem, and sometimes even beating the 2500k.

While sure it's not superdooperawesome like a lot of people were hoping for, but it's far from fail IMO.

EDIT: W1zzard, how come you guys don't test in 1080p? I'd dare say that the probably THE MOST common resolution used these days, yet it's always omitted in tests.

That is because at higher resolutions, your bottleneck is almost always your video card. Bulldozer didn't fail, it just had a lot of hype for something that was decent at best in comparison to the T1100 and T1090. Bulldozer did good enough where it has to and shines when SMP really matters. It's the first step towards something better. There aren't a whole lot of applications that use a lot of SMP, but there very well could in the future.

This review kind of puts everything into perspective imho.
http://guru3d.com/article/radeon-hd-7970-cpu-scaling-performance-review

In all realism, BD isn't that bad. Keep in mind that a lot of these titles don't use a lot of cores yet, so there is a lot of horse power BD still has waiting to be used. Nothing is stopping someone from transcoding video while playing a video game and not have a problem. That is what AMD is trying to do.

Albeit, Intel has better IPC counts, but that is only because SB has a shorter pipeline than BD. BD has some obstacles to overcome, but all in all, it is more space friendly, so you can cram more cores on the same amount of die space.

 

[BlackOmega]

That is because at higher resolutions, your bottleneck is almost always your video card.

I know that, but thanks anyway. The reason I'm asking is because why would you review at a resolution that 7% use, as opposed to a resolution that is THE MOST commonly used --1920x1080 (25%). Which is followed by 1680x1050 @ 17%. (source)

Bulldozer didn't fail, it just had a lot of hype for something that was decent at best in comparison to the T1100 and T1090. Bulldozer did good enough where it has to and shines when SMP really matters. It's the first step towards something better. There aren't a whole lot of applications that use a lot of SMP, but there very well could in the future.

This review kind of puts everything into perspective imho.
http://guru3d.com/article/radeon-hd-7970-cpu-scaling-performance-review

In all realism, BD isn't that bad. Keep in mind that a lot of these titles don't use a lot of cores yet, so there is a lot of horse power BD still has waiting to be used. Nothing is stopping someone from transcoding video while playing a video game and not have a problem. That is what AMD is trying to do.

Albeit, Intel has better IPC counts, but that is only because SB has a shorter pipeline than BD. BD has some obstacles to overcome, but all in all, it is more space friendly, so you can cram more cores on the same amount of die space.

You're absolutely right, it didn't but all of these intel fanboi's would lead people to believe that it's slower than socket 939 single core. When in reality, it's their fastest CPUs' to date. Regardless of whether the software can utilize it or not.

 

[cadaveca]

Regardless of whether the software can utilize it or not.

For most peopel though, this is what's most important.

The question gets asked "Will this make what I do now faster?"

And the answer, of course, for most is "Not Really".


The hype let people down, but of course it's not a bad chip...but it's not "The best" either. I think many more people would be happier if the 8150's price matched the 2500k's, but it doesn't.

I've recommended to many PHenom II quad users that they upgrade to 8150, and soem have. Not one has been disappointed in the change.

 

[ensabrenoir]

I know that, but thanks anyway. The reason I'm asking is because why would you review at a resolution that 7% use, as opposed to a resolution that is THE MOST commonly used --1920x1080 (25%). Which is followed by 1680x1050 @ 17%. (source)

You're absolutely right, it didn't but all of these intel fanboi's would lead people to believe that it's slower than socket 939 single core. When in reality, it's their fastest CPUs' to date. Regardless of whether the software can utilize it or not.

Aaaaahhhhhhhh thats like having the fastest speed boat....in the middle of the desert. JUST ENJOY THE DARN THING!!!!!!!! There is no mighty, morphing power ranger upgrade fix for it. Its not the intel killer. By the time software catches up to it....sometime 10 times better will be available... I still use my mindisc and laserdisc player....why? Because I paid for it and it makes me happy....get it? Apply

 

[NdMk2o1o]

Don't really see what all the fuss is about with BD, it compares and competes directly with Intel Nehalem which before SB was the target. I would be happy with BD if I had one, just so happens I have SB and am able to BD lol but seriously, still a half decent chip and if anything with AMD the 2nd revision will always be stronger as they improve upon the 1st build. Nice review W1zz, thanks as always!!

 

[Aquinus]

You also have to realize that single-threaded workloads isn't bulldozer's strong suit. You do some media encoding with the 8150 and it will give any 1155 CPU (at the moment,) a run for it's money, and in some cases get's close to 990X performance when it comes to media. If I look at framerates for any game and see them practically at 50-60fps, I wouldn't complain. Also SC2 isnt as dependent on IPC as it is on memory bandwidth, which is what Intel's chips are currently excelling it. Just keep in mind that a properly tuned bulldozer can crank out some impressive numbers.

Also, rumor has it that the next version of bulldozer, "Enhanced Bulldozer," may have a quad-channel memory controller while still using AM3+. That could be a good selling point as AMD processors are reasonable to replace without having to change all of your hardware. Intel's IPC counts are much nicer than AMD's, but AMD has something going for it because bulldozer has a very scalable architecture. Once AMD trims off the fat, reduces the length of the pipeline and gets its memory controller up to snuff. It will do better in single-threaded applications, and there will be more cores at the same time.

The future isn't single-threaded applications, just keep that in mind. Remember where we were 10 years ago, and 10 years ago before that.

 

[Prima.Vera]

Damn. I have a 4 years old Core 2 Quad Q9650 which is on pair with the I7-920, and still beats the crap out of the Shitdozer. Shame AMD, shame!

 

[Mussels]

Damn. I have a 4 years old Core 2 Quad Q9650 which is on pair with the I7-920, and still beats the crap out of the Shitdozer. Shame AMD, shame!

show me where you got these numbers from, please

 

[InnocentCriminal]

Yeah I'll 2nd that. I'll be interested to see how it compares with my Q9550.