[TomsHardware] CrossFire Scales Spectacularly

[Annisman*]

http://www.tomshardware.com/reviews/crossfire-sli-3-way-scaling,2865-12.html

From Tom's Hardware:


"CrossFire came out with a huge overall scaling lead over SLI, and removing the one title that didn’t reflect that average would have made the lead even bigger. Superior scaling allowed two mid-priced Radeon HD 6950s to approximate the performance of two higher-cost GeForce GTX 570s, while three HD 6950s took the performance win over three GTX 570s. "

How are we going to explain this.... ?

Can we stop saying that ATI drivers are 'crap' now ? And that CF is just as good if not much better than SLI at the moment.

Now I hope Nvidia sees this and gets their act together on scaling, because I've got a second GTX 580 on the way


Intentionally inflammatory/baiting thread-titles are not going to work. The added rhetoric and hyperbole is needless.

Also, per thread citation requirements, unless the data comes from yourself you need to cite the source of the results being used to justify the thread title.

Original Thread Title: "Uh Oh! Get the popcorn, Nvidia just got wasted in multi-card scaling"

New Thread Title: "[TomsHardware] CrossFire Scales Spectacularly"

Moderator Idontcare

 

[Powermoloch]

nice results :thumbsup:

 

[badb0y]

"GeForce GraphicsGeForce/Ion 266.58Radeon GraphicsAMD Catalyst 11.1"
Whats the newest drivers for nVidia? I think they fixed some of the scaling issues with their new drivers.

http://alienbabeltech.com/main/?p=24236

Another great review of CF vs SLi, seems they used the same drivers.

 

[OCGuy]

Could this thread possibly end up requiring popcorn in part due to your carefully chosen thread title?

CF was terrible on the 4XXX series, I am glad to see it is getting better. I would still see no reason to ever go over 2 GPUs...much too inefficient no matter what brand you choose.

 

[santz]

Could this thread possibly end up requiring popcorn in part due to your carefully chosen thread title?

CF was terrible on the 4XXX series, I am glad to see it is getting better. I would still see no reason to ever go over 2 GPUs...much too inefficient no matter what brand you choose.

THIS, (I love saying that)

 

[4ghz]

Any one else think its kind of ridiculous and irresponsible for some of these websites to draw these giant, sweeping conclusions after testing a measly 5 games? Hell I've seen it done with as little as 3 games. *cough* hardocp *cough*.

These results don't really matter to me at the moment as I currently have a 5870 2gb and I don't plan to crossfire because I assume the 5 series scaling still sucks. But its kind of shocking how some of these "respected" and often quoted sites can be so lacking in judgment. They simply spit out these incomplete articles to chase some quick hits.

 

[Aristotelian]

Any one else think its kind of ridiculous and irresponsible for some of these websites to draw these giant, sweeping conclusions after testing a measly 5 games? Hell I've seen it done with as little as 3 games. *cough* hardocp *cough*.

Indeed, and I've been posting comments like that for over 6 months now.

 

[notty22]

They tested 5 games and wanted to throw 1 out ? Here is a better article with 18 games.

\
\

Whats up with title ?

edit: I see , someone commented the same, about 5 games. I was reading TH
to be fair.

 

[Chaoticlusts]

I'm always a bit dubious about results from toms (not that they're never right but still prefer other sources)...used to read the website heaps but several years ago it really went downhill...numerous errors in every review and somethings that just seemed flat out bias...found anand and was much happier ...dunno if they've improved since then mebbe they have but still generally prefer to get news like this from other websites

(I own a 6950 so I'm not saying this out of nvidia bias or anything)

 

[Lonyo]

Most websites from the HD6800 and 6900 launch showed that both had superior dual card scaling than NV cards.
Someone even did a comprehensive roundup in a google docs spreadsheet showing the average dual card scaling across many games to be something like 80% for AMD and 70% for NV or something similar to that.

 

[1h4x4s3x]

It is really puzzling that reviewers, at this day and age, don't own 3 monitors.

 

[Red Storm]

Could this thread possibly end up requiring popcorn in part due to your carefully chosen thread title?

CF was terrible on the 4XXX series, I am glad to see it is getting better. I would still see no reason to ever go over 2 GPUs...much too inefficient no matter what brand you choose.

Err...

i7 950 @ 3.9ghz
2x XFX 5850s Xfire

Edit: Ah, you said over two GPUs.

 

[nOOky]

lol yea, like OC Guy I see no reason to ever have two cards either!

 

[wahdangun]

They tested 5 games and wanted to throw 1 out ? Here is a better article with 18 games.

\
\

Whats up with title ?

edit: I see , someone commented the same, about 5 games. I was reading TH
to be fair.

the problem is, why waste money for just 1080P monitor ????

here are real men resolution

nv really need to step up their Vram to be competitive with AMD or maybe its nvdia strategy so it will never eat GTX 580 market (because GTX 570 was easy to oc to GTX 580 performance)

 

[Vdubchaos]

Could this thread possibly end up requiring popcorn in part due to your carefully chosen thread title?

CF was terrible on the 4XXX series, I am glad to see it is getting better. I would still see no reason to ever go over 2 GPUs...much too inefficient no matter what brand you choose.

Agreed
I will never go/plan for Dual GPU set up.

1 card only please...

 

[Idontcare]

I'm always a bit dubious about results from toms (not that they're never right but still prefer other sources)...used to read the website heaps but several years ago it really went downhill...numerous errors in every review and somethings that just seemed flat out bias...

The old tom's died when Pabst sold out to the Media Marketing Group (I think that was name of the initial buyers). After that it just became a review-churning money making operation, quality went down, page sizes got smaller and smaller so they could get 20 page impressions from a minor review instead of the 2 or 3 pages it should have been, and the reviewers themselves weren't exactly cut from the same cloth as Tom.

Anand is one the last of the original old-school independently owned review sites. There are new ones like ABT, and maybe [H] would count as new and not old-school, I don't remember them from the K6/pentium days like Anand and Tom's, but there's no school like the old-school IMO.

 

[WhoBeDaPlaya]

Anand is one the last of the original old-school independently owned review sites. There are new ones like ABT, and maybe [H] would count as new and not old-school, I don't remember them from the K6/pentium days like Anand and Tom's, but there's no school like the old-school IMO.

Man, I remember the days of Tom's Overclockers mailing list. We forked off and spawned The Hardware Group from that.

 

[maddie]

Could this thread possibly end up requiring popcorn in part due to your carefully chosen thread title?

CF was terrible on the 4XXX series, I am glad to see it is getting better. I would still see no reason to ever go over 2 GPUs...much too inefficient no matter what brand you choose.

Inefficient by what parameter?


With crossfire scaling like this, do you realize that its possible to get greater performance from a given # of shaders using 2 cards rather than having a bigger GPU.

If they can get it working automatically in more games, then we might have a multi-GPU future.

 

[tviceman]

I'd like to know what rig they were using. Even at 2560x1600 resolution, a core 2 quad can be holding back a SLI / Crossfire configuration. I went from a c2d e8400 @ 4.05ghz to a sandy bridge and got huge gains even in games that didn't benefit from 4 processors. Also, there was an article a few weeks ago on here that grooveriding posted showing that Fermi needs more CPU horsepower to maximize it's full potential.

System specs should always be posted when benchmarking is done.

 

[busydude]

System specs should always be posted when benchmarking is done.

http://www.tomshardware.com/reviews/crossfire-sli-3-way-scaling,2865-2.html

 

[tviceman]

http://www.tomshardware.com/reviews/crossfire-sli-3-way-scaling,2865-2.html

Ah thank you.

 

[Will Robinson]

Just accept that AMD has superior multi card scaling.
No need to kill/diss the messenger.^_^

 

[pcm81]

Here is the lo-down on scaling:
scaling applies to resources, not just processing power. First couple words about how scaling works:
In GPU codebase the SIMM code model is used, which stands for Single instruction, multiple memory. The massively parrallel computing tasks from Boinc client or for that matter any OpenCL like application uses SIMM model. Lets say you are doing binary search on anarray of integers:
1. in the case of single processor (1 core) you divide the array into 2 partch, check if element you are serching for is in the bottom part and then repeat this checkb by dividing bottom or top part into 2 smaller parts. thisis basic binary search algorithm. if your total array is 1024*1024 elements in size it will take 20 iterations of this algorithm to find the number you are looking for.
2. now imagine that you have 1024 processors all able to access the array in ram at the same time, but in different locations. You divide the array into 1024 parrts and assign each core to check 1 chunck to see if searched for element is in that chunk. On second iteration you break up the chunk found in 1st iteration and break it up into 1024 more elements. So after 2nd iteration you have the answer. In this case we see 10x speed improvement over a single core.

Now, lets say you have 6000 cores (2 radeon 6990s in quad fire). You can render 6000 regions of the screen in one iteration of rendering algorithm. Each region has a minimum allowable size, because below this minimum size the overhead of task assigning to the cores is too big for the small amount of time it takes 1 core (stream proccessor) to render the chunk of the screen assigned to it. The length of 1 iteration of the rendering algorithm increases with detial level of the game. So, what you have here is 2 factors: speed of 1 core and total number of cores. I suspect that Nvidia cores are faster, because at lower resolutions nvidia gpus win. However at higher resolutions the Nvidia GPU with fewer cores has to run the rendering algorithm more times than ATI GPUs have to run it, because each iteraion of ATI's GPU renders larger section of the screen.

Between the iterations of this rendering algorithm sinchronization must occur which is essentially overhead which is proportional to total number of cores (Cuda cores or stream processors). So what we have is a longer overhead time for a system with 6000 stream processors than for a system with 3000 cores... This is why you see a single 6990 beat 2x6990 at lower resolutions. At higher resolutions, when a single 6990 has to repeat the rendering algorithm more times than 2x6990 the quadfire wins. The same analogy applies to Nvidia cards vs radeons.

 

[gevorg]

Here is a better article with 18 games.

Nice! xbitlabs > toms

 

[cusideabelincoln]

I'd like to know what rig they were using. Even at 2560x1600 resolution, a core 2 quad can be holding back a SLI / Crossfire configuration. I went from a c2d e8400 @ 4.05ghz to a sandy bridge and got huge gains even in games that didn't benefit from 4 processors. Also, there was an article a few weeks ago on here that grooveriding posted showing that Fermi needs more CPU horsepower to maximize it's full potential.

System specs should always be posted when benchmarking is done.

Crossfire and SLI are more dependent on quad cores (and CPU power) than single cards are. You probably would have seen the same kind (trend) of gains if you moved to a C2Q @ 4 GHz.

Besides your conclusion isn't very conclusive here. Sandy Bridge is two generations ahead of Wolfdale. It has higher cores and also higher IPC, and can reach higher clockspeeds. You could have simply been CPU limited, not necessarily core limited, in your games. Although it's likely you were both.