CrossfireX vs SLIQuad

As an overclocking junkie I would like to try and clear up an issue which has been bugging me for a while: in a 4 GPU arrangement, is it better to side with ATI or nVidia? :help:

All of the top scores on hwbot.org (sorry cant post URLs yet, :angst:) for 3DMark06 are achieved using 4 Radeon 4890 cards in CrossfireX combined with an overclocked Xeon W3540 or i7 975 CPU.

I can fuly understand the choice of CPUs - them both being the best performing CPUs around with at least 4 cores and huge overclocking potential. But the GPU choice confuses me because a Radeon 4890 is less powerful than a 285GTX (i.e. 1 half of a 295GTX). So when you combine 4 of each surely the CrossfireX 4890s have less total power than a pair of 295GTX cards?

If you are still with me, can you thus explain why the paired 295GTX cards are less powerful! Is it a case of the performance of the cards from nVidia scaling less well than those of ATi when combined in multi-card setups? (I have heard about nVidia performancedecreasing relatively as more cards are added in a setup, i.e.: 1X, 2X, 2.8X, 3.4X. But is the same scaling issue true for ATI based cards?)

If anybody is particularly keen to rectify my muddled mind, does somebody know if combining 2 ASUS MARS 295GTX would produce a significant performance boost over 2 standard 295GTX cards in SLI; and whether the ASUS combo would beat 4 4890 cards in CrossfireX?

I most heartily thank anybody who can clear this issue up for me! :bowdown: Although my piggy bank is not quite up to paying for any of the aforementioned setups to do a personal comparison, I would be grateful for the knowledge. Thanks again! :hexlub:

From what I know, and I fully expect to be ripped to shreds

Tri-SLI is the max, though if its 2 dual gpu cards I really dont know what will happen

Due to scaling of both SLI and x-fire I doubt the difference would be hugely noticeable

Thanks Fuzz,
If you put 2 dual GPU cards together in SLI you get quad SLI.

Its the sort of notion that "I doubt the difference would be hugely noticable" which I would like to clear up, because there is a 3000 Mark gap in hwbot.org 's records for 3DMark06 between 2 295GTX cards in SLI and 4 4890 cards in x-fire.

3000??? seriously?

I'm going to look that up, would never have expected it to be anywhere close to that much and it sounds a good read if there is an article to boot.

Sorry for the apparently extremely false information.

No problems,
It is this huge gap between the 38000 marks of the 4890s and the 35000 marks of the 295GTXs which has left me asking why so if the 4890 is the less powerful card!

http://www.tomshardware.com/reviews/...-295,2123.html

Did you crossreference this article with another? or is there one with both the 4890 Quad-Fired and GTX295 SLI as this one only has a double helping of 4870x2 to compare nVidia's pair of monsters to, even so at high resolutions on crysis (2560*1600) I was surprised to see the 4870x2 crossfired beating the GTX295s with an underdog-stick into the ground

Nobody seems to have noticed, a GTX295 is not two GTX285s, it is two GTX260s.

Quote:

---

Originally Posted by sammorris

Nobody seems to have noticed, a GTX295 is not two GTX285s, it is two GTX260s.

---

But i thought it was 2 GTX 275's :confused: and becuase of this the reason everyone was fapping over them new Asus MARS cards was because they combined 2x GTX 285 rather than the reference 2x 275's.

Also, dont forget the memory speed those 4890's push and 3Dmark doesnt really show how they perform game wise.

The problem I have noted with beyond SLI and Crossfire configurations that include in excess of 2 cards is that of yet, the balance is not very effective. With 2 cards the scaling is almost linear, but it gets progressively worse with the more cards you add.

Nope, it's two 260s. If it were 275, the GTX295 would actually be worth its absurdly high asking price.
Nightkhaos: Recent drivers have changed that. There are quite a lot of games showing over 300% scaling for 4 GPUs.

Including some rather large claims in the very latest ATI/AMD driver release notes.

One of them at least (Crysis Warhead) is true. The performance difference there is vast.

Quote:

---

Originally Posted by laywill

As an overclocking junkie I would like to try and clear up an issue which has been bugging me for a while: in a 4 GPU arrangement, is it better to side with ATI or nVidia? :help:

All of the top scores on hwbot.org (sorry cant post URLs yet, :angst:) for 3DMark06 are achieved using 4 Radeon 4890 cards in CrossfireX combined with an overclocked Xeon W3540 or i7 975 CPU.

---

Are you comparing scores with the same overclocked CPU? Multi-GPU setups are more or less entirely constrained by CPU above 3 GPUs. When that happens then additional cards actually reduce performance because of increased CPU overhead.

Quote:

---

If you are still with me, can you thus explain why the paired 295GTX cards are less powerful!

---

If the CPUs are exactly the same yet 2x295 is slower then it's simply that the CPU overhead is more, and both are constrained by CPU.

Quote:

---

If anybody is particularly keen to rectify my muddled mind, does somebody know if combining 2 ASUS MARS 295GTX would produce a significant performance boost over 2 standard 295GTX cards in SLI; and whether the ASUS combo would beat 4 4890 cards in CrossfireX?

---

No in both cases, due to the CPU limitations. Such limitations are being addressed as much as possible now, especially in conjunction with chipsets (AMD have started concentrating on whole system performance, so how to optimise communication between cards, chipset and CPU for example, which should show some benefits. However you can't seriously consider AMD systems at the very top level unfortunately due to the dominance of Intel high end CPUs).

Depends on the 3dmark, 3dmark06 is an absolute CPU bottleneck at 1 GPU, but Vantage shows a reasonable improvement with multiple cards.
Four 4890s are actually a more powerful combination than two 295s, as you're basically comparing 4 GTX260s vs 4 HD4890s - and the HD4890 is a much more powerful GPU. In low-resolution environments where the GTX295's poor memory performance doesn't show up, it will beat two HD4870X2s because Quad SLI scaling is better, but not four HD4890s.

Two MARS cards would be better, since the GTX285 is more powerful than the HD4890 (though not by very much, considering this is £2050 vs £530, maybe 10-15% at most?). If you're going for quad crossfire, as said, AMD CPUs are slightly more optimised, but their lack of performance means at this level, Intel are your only option. The new driver performance updates apply to Intel as much as AMD from what I can see.

Quote:

---

Originally Posted by sammorris

Four 4890s are actually a more powerful combination than two 295s, as you're basically comparing 4 GTX260s vs 4 HD4890s - and the HD4890 is a much more powerful GPU.

---

The GPUs in the 295GTX have more shaders than the 260 - the same number as the the GTX275 in fact. Though it has a lower clockspeed than the latter.

But as I mentioned, talk of GPU power is meaningless when you're bottlenecked by the CPU/subsystem.

You certainly are in 3dmark, but in real world applications, the CPU bottlenecks aren't always as apparent at high resolutions. Let's face it, to be considering quad SLI/CF, you should really be using 2560x1600, 1920x1200 perhaps, but dual graphics can deal with 1920 pretty well enough (especially two 4890s)

Thanks to everyone!

I was totally unaware of the fact that a 295 is a set of paired 260GTX cards! That explains why an i7 with 4 HD4890 cards blow the i7s paired with 295GTX cards out of the water.

I had a look at the top results for 3DMark Vantage on hwbot.org here. The top two results use both different CPUs and different GPUs. The holder of the highest score has an i7 920 @ 4GHz on water cooling with a pair of 295GTX cards and scored 32270 marks. Second place used a Xeon W3570 @ 5.2GHz on Ln2 combined with 4 HD4890s.

If, as has been suggested earlier, the 295s in SLI are less powerful than 4 HD4890s in x-fire, and that most benchmarks with multiple GPUs end up CPU bound, why has a 4GHz i7 with nVidia GFX triumphed over the higher clocked Xeon with the 4890s?

I know that AMD have been pushing the dragon platform (Phenom + 790GX mobo + HD4870) and are soon to be releasing an updated version (codename Leo?) with support for higher clocked DDR3 RAM or something along those lines. However, Phenoms do not compare to an i7 or a Xeon so are totally useless for a record attempt.

Have I just gone and confused the whole issue here? As far as I now see it: 295GTX cards are great in single card setups; for multi GPU setups go for HD4890s because of higher performance and a big saving for the piggy bank. So is there a huge difference, which is affecting the performance results, between an i7 and a Xeon W3570?

Thanks again for everybody's help o far in sorting a muddled mind :bowdown:

Yeah, the GPUs are slightly modified, but GTX275s they ain't (the 295 came out before the 275 I think) - the other issue is that in high memory environments (mostly 2560x1600) the GTX295's 896MB per GPU of memory coupled with nvidia's weaker memory management means AA performance is very unpredictable. Four 1GB HD4890s are a more powerful and sustainable combination, IF you can find a motherboard that will fit four dual slot cards, they are very few and far between. The only board I've found designed straight-up for the job is an old MSI Phenom board.

I wouldn't worry yourself with 3dmarks. The HD4870 and GTX260 are rough equals yet the GTX260 absolutely demolishes the 4870 in 3dmark, remember nvidia have about 10 years of driver coding to abuse 3dmark for unrealistically high scores. Look at real-world benchmarks, since you aren't going to be playing 3dmark!

To be honest, I never advocated GTX295s. A pair of 4890s, or even a 4870X2 were a better buy IMO due to the lower price, and more consistent performance.

Quote:

---

Originally Posted by laywill

I was totally unaware of the fact that a 295 is a set of paired 260GTX cards!

---

They're not - they're paired 275s, which when released as single cards could be clocked higher.

Quote:

---

Have I just gone and confused the whole issue here? As far as I now see it: 295GTX cards are great in single card setups; for multi GPU setups go for HD4890s because of higher performance and a big saving for the piggy bank. So is there a huge difference, which is affecting the performance results, between an i7 and a Xeon W3570?

---

Natively? Not all that much, but when overclocked I don't think the only thing they are adjusting is CPU frequency..

If you want a pair of GPUs then get a single 295. If you want triple GPUs then get 3x4890. Anything else is going to be CPU or subsystem limited unless you take extreme measures to work around those.

Thanks for the advice!
Cheers everybody.

Quote:

---

Originally Posted by sammorris

Yeah, the GPUs are slightly modified, but GTX275s they ain't (the 295 came out before the 275 I think)

---

It has the same amount of SP's per GPU as the 275 (240) albeit it clocked at the 260 references, so id still say it has more in common to the 275 than the 260 :juggle:

Edit: Would dual-socket motherboards help relieve the bottleneck ?

Quote:

---

Originally Posted by Terbinator

Edit: Would dual-socket motherboards help relieve the bottleneck ?

---

Not by themselves.

You need:
1) An application that has an efficient and multi-CPU and GPU core scaleable graphics rendering pipeline.
2) An API that is efficient and multi- CPU and GPU core scaleable.
3) Non-limiting connections from GPU-GPU, CPU-CPU and GPU-CPU.
4) Non-limiting CPUs.
5) Non-limited GPUs.

Drivers have only recently started becoming multi-CPU core aware. DirectX11 will bring some awareness to the windows API.

I'm pretty certain they're paired 260s, because two 275s in SLI give a 295 a noticeable beating. If you want a pair of GPUs, buy a pair of HD4890s. if you want three GPUs, buy three 4890s. I think you can see where this is going. Were GTX275s reasonably priced, I would recommend them instead. The ridiculous cost of a GTX295 means it is never going to be worth it in the face of the 4890s.
This "bottleneck" only really applies to 3dmark. In real world applications you're not going to see it as often unless you're using pointlessly low detail for the graphics setup you're using. (For ref, 3dmark will only see the extra cores of a dual CPU system in the CPU tests, the graphics tests will be unaffected).

Quote:

---

Originally Posted by sammorris

I'm pretty certain they're paired 260s, because two 275s in SLI give a 295 a noticeable beating.

---

Even if you clock the 295 chips up to an equivalent clockspeed?

The thing is though the HD4890 and HD4870X2 are just better value for most people ATM.However the GTX275 is starting to drop in price nearer to the overclocked HD4890 cards though. Crossfire is better supported on BOTH Intel and AMD platforms and the motherboards tend to be cheaper too. However it seems that HD4870X2 cards are in very short supply now - perhaps production has been wound down in anticipation of the new DX11 parts??

TBH,a single reference clocked HD4890 will run most current games at decent setting at 1680x1050 and even 1920x1200 which is not bad for a graphics card under £150.

kalniel: Even if you clock the 295 up to an equivalent clock speed, two 275s still beat it, sometimes by up to 10%.
Cat: The HD4870X2 is unfortunately irrelevant as it no longer exists (was discontinued at source by ATI at beginning of July), HD4890s are all we have to play with until the HD5870s come out for the moment.

When DX11 gpu's come out there will be a dual GPU...
lets see how that compares. I've heard stories that quad ATIs see terrible performance in games but who knows, depends on the game I guess.

As someone who uses Quad CF I assure you it's not the case. trouble is, the system was invented before the drivers for it were any good, so most reviews were published back when it genuinely was rubbish. Look at any recent reviews and it fares much better.
Historically, the dual card for a generation does not come out as early as the solo counterpart so we'll have to see what happens. Of course with the HD4870 it was only a month after that the X2 arrived. Here's hoping the same applies to the HD5870.

Quote:

---

Originally Posted by sammorris

As someone who uses Quad CF I assure you it's not the case. trouble is, the system was invented before the drivers for it were any good, so most reviews were published back when it genuinely was rubbish. Look at any recent reviews and it fares much better.
Historically, the dual card for a generation does not come out as early as the solo counterpart so we'll have to see what happens. Of course with the HD4870 it was only a month after that the X2 arrived. Here's hoping the same applies to the HD5870.

---

Interesting. Nice system you got there then :D
So it will be the 5 series and carry on with the 58 55 57 series etc like the 4000 series?

I had a rumour that the new ATI cards were going to be called the "7 series" although this could be not true.

As far as I can see they will have the same sort of system, but the HD5870 name has not come from ATI themselves, there is no official name. HD5870 is the most likely, but there is no concrete evidence to support that yet, a little unusual given the product is only 3 weeks away.

Quote:

---

Originally Posted by sammorris

As far as I can see they will have the same sort of system, but the HD5870 name has not come from ATI themselves, there is no official name. HD5870 is the most likely, but there is no concrete evidence to support that yet, a little unusual given the product is only 3 weeks away.

---

I reckon it's gonna muck up my plans to buy a 4890 in december even more :P

Nah I suspect you'll be buying something other than a 4890.

Quote:

---

Originally Posted by sammorris

Nah I suspect you'll be buying something other than a 4890.

---

The problem is that it will obviously be better to get the newest card so I am at least up to date for a year or so, but it won't necessarily mean I'll get more performace for the same price bracket. I'll probably have to form out £200 which I don't really want to...

That depends how much the new cards cost. The 4870s didn't cost as much as £200 even when they first came out.

That's the 512mb version though. I reckon if I want 1GB then the 4890 will offer better performance at the £150 price point. By christmas though, prices might have fallen considerably!
It could work out perfectly...

True but I would imagine that, like the 4890, the standard will be 1GB. I wouldn't have thought they'd start with a card as powerful as the (5870) and give it only 512MB of memory, but who knows.
Personally I'm rather hoping for a 4GB (2GB per GPU) X2 card rather than the standard 2GB one. 2GB is usually enough but for some games at 2560x1600 maxed (which is the reason I'll be upgrading) 2GB doesn't quite cut it, especially with AA (Or for GTA4 :S)

Would be upward of £450 for 4gb you'd have to guess. I'll leave that one to you then :D

I won't be able to afford any replacement hardware until mid-october, so I'll see what the situation is by then...

You could think about getting a few of these 4870s with 2Gb memory and put them in x-fire...

Quote:

---

Originally Posted by laywill

You could think about getting a few of these 4870s with 2Gb memory and put them in x-fire...

---

Two 4870 2GB in crossfire would not be better than his current 4870x2 2GB in crossfire.

As for this whole thread. You can't take hwbot top scores into account to compare like for like with ati and nvidia cards, there are some very good tweaks out there that can make all the difference when you are at the very limit.

with the 4890 at £126 not 150, your statement is wrong. For £150 a newer card will be avaliable.

The R800 promises almost double the performance of the previous, so the £100 card (probably 5850) will be like a 4890 or more?

wow!
That will make the top of the range cards phenomenally quick then!

Peter, surely adding an extra 4GB of memory buffer would be beneficial? Or is that overkill and you're better buying a card with more horsepower?

Quote:

---

Originally Posted by laywill

wow!
That will make the top of the range cards phenomenally quick then!

Peter, surely adding an extra 4GB of memory buffer would be beneficial? Or is that overkill and you're better buying a card with more horsepower?

---

Unless youve got some uber h4x monitor with a resolution of 9999*8888 4GB is pointless.

All the tests I saw found no benefit to having the additional memory on those cards... no card out there can move quickly enough to use it

wellt eh GTX275 is made by being based of a GTX295, so id say that the GTX295 is 2 downclocked 275s.

yes 275 in SLI si better, but the actualy GPU it self is the same.