ATi 4870 X2 (R700) thread

[Janooo]

Current info

AMD bests Nvidia with graphics chip strategy
Pictures #1
Pictures #2
PLX PCI Express switches
4870 X2 15% better than 4870 Crossfire?
R700 Crossfire Vantage score
4870 X2 readies for release
4870 X2 clock speeds
2 x 1 GB memory
R700 is also Radeon 4850 X2 2GB
R700 up to 80 % Faster than GeForce GTX 280
R700 GDDR5 to sell for $549?
Qimonda to ramp up GDDR5 production
Leaked ATi slides (August 12 release?)
Interview with ATi (confirms August launch).
R700 preview @driverheaven.




Thread has been replaced with reviews thread.

Video Mod BFG10K.

 

[unr3al]

Well done AMD/ATi! Now all they need is a decent CPU to replace Phenom/AthlonX2.

 

[chizow]

Its a very impressive looking piece of hardware, but I think there's going to be some things prospective buyers need to look for and be concerned about:

1) Still using a PLX bridge. Basically splits the PCIE x16 lanes for 2 x RV770 GPU, resulting in 2x PCIE x8. As we saw in the Tweaktown review, this can be a problem as PCI x8 2.0 does show bottlenecking with 4850s. This will be an even bigger problem for PCIE 1.1 board owners, as you will only have an effective PCIE x4 2.0 for each GPU.

2) CPU bottlenecking. Make sure your CPU is fast enough to take advantage of the extra GPU at whatever resolution you play at. 4870 doesn't scale that well in CF due to CPU bottlenecking in many games so you may not see any benefit until faster CPUs are available.

ATI has shown excellent scaling in games that scale well with the 4850s, so the CF Sideport can only improve things there. Still need to see if it helps micro-stutter as they claim, or if it allows them to share a common memory pool.

 

[AzN]

We still haven't seen if R700 implementation yet but current X-fire still needs quite a bit of work to be good as a single GPU or even match SLI scaling or functionality. When ATI x-fire does work it scales good but majority games fail even the new RV770. I think this article is propaganda/fanboy article at best. ATI needs to prove this first like RV770 did for price performance. So far they haven't shown anything.

 

[MyLeftNut]

Are there actually tests done with a 4850 in pci-e 2.0 x8 in single card which show the bottleneck? I've seen crossfire tests on x48 vs p45 that show a deficit, but then who knows what the exact cause is then. On the other hand, I've seen single card tests with a 8800gts (g92) which show it to be on par on the same amount of bandwidth as a pci-e 2.0 x8 (in this test they used pci-e 1.0a x16).

http://www.tomshardware.com/re...ci-express,1761-7.html

Too bad they did not do the same comparisons with something more powerful at that time like an 8800 ultra. Also knowing that the 4850 doesn't perform that much better than an 8800gts g92 may perhaps indicate the issue with underperforming 4850s in pci-e 2.0 x8 isn't in the available bandwidth.

 

[thilanliyan]

Originally posted by: chizow
Its a very impressive looking piece of hardware, but I think there's going to be some things prospective buyers need to look for and be concerned about:

1) Still using a PLX bridge. Basically splits the PCIE x16 lanes for 2 x RV770 GPU, resulting in 2x PCIE x8. As we saw in the Tweaktown review, this can be a problem as PCI x8 2.0 does show bottlenecking with 4850s. This will be an even bigger problem for PCIE 1.1 board owners, as you will only have an effective PCIE x4 2.0 for each GPU.

How do you know they'll use the same PLX bridge chip as the 3870x2?

According to a thread I've seen at Xtremesystems, and from these pics (IIRC the previous one was a larger chip) (http://en.expreview.com/2008/0...al-r700-pics-unveiled/), the chip they use is different and so won't necessarily be x8/x8. And in all likelihood, if there was a bottleneck in that part of the card from the 3870x2, ATI would have replaced it this time around.

You're talking as if what you say is fact when most people don't know much about the card to begin with. If you know otherwise then please feel free to share or at least state that you have access to privileged information. Otherwise, just search around a bit instead of jumping to false conclusions.

 

[Lonyo]

Wow, that article was atrocious in terms of its inaccuracies.

 

[SlowSpyder]

Originally posted by: chizow
Its a very impressive looking piece of hardware, but I think there's going to be some things prospective buyers need to look for and be concerned about:

1) Still using a PLX bridge. Basically splits the PCIE x16 lanes for 2 x RV770 GPU, resulting in 2x PCIE x8. As we saw in the Tweaktown review, this can be a problem as PCI x8 2.0 does show bottlenecking with 4850s. This will be an even bigger problem for PCIE 1.1 board owners, as you will only have an effective PCIE x4 2.0 for each GPU.

2) CPU bottlenecking. Make sure your CPU is fast enough to take advantage of the extra GPU at whatever resolution you play at. 4870 doesn't scale that well in CF due to CPU bottlenecking in many games so you may not see any benefit until faster CPUs are available.

ATI has shown excellent scaling in games that scale well with the 4850s, so the CF Sideport can only improve things there. Still need to see if it helps micro-stutter as they claim, or if it allows them to share a common memory pool.

I don't really buy #2... I didn't read everything you and BFG wrote back and forth, so I didn't see if you had any examples of CPU bottlenecking. I could be wrong, and I'd be interested to see some numbers, but I just don't think a modern CPU will bottle neck video cards at any normal-high resolution/high quality settings. It's possible there was a bottleneck somewhere in a CF system that held it back, but I just don't think that it would be a CPU bottleneck. Infact I'm willing to bet my Phenom would put out very similar numbers as an overclocked C2D system would with a given videocard on a 24" monitor.

 

[Foxery]

Originally posted by: Lonyo
Wow, that article was atrocious in terms of its inaccuracies.

Care to elaborate? C|Net is the Reader's Digest version of a real tech site like AnandTech, but they're quite well respected and credible.

The only error I saw was the figure for "possible scaling to 4 GPUs." 1.8*1.8 = 3.24. The 2.5 figure seems plucked out of thin air.

 

[chizow]

Originally posted by: thilan29
the chip they use is different and so won't necessarily be x8/x8. And in all likelihood, if there was a bottleneck in that part of the card from the 3870x2, ATI would have replaced it this time around.

What do you think the bridge chip does? Unless the PLX gen 2 chip is able to make an x16 PCIE 2.0 slot into an x32 PCIE 2.0 slot then it will result in an x8/x8 PCIE 2.0 on a single slot.

 

[chizow]

I don't really buy #2... I didn't read everything you and BFG wrote back and forth, so I didn't see if you had any examples of CPU bottlenecking. I could be wrong, and I'd be interested to see some numbers, but I just don't think a modern CPU will bottle neck video cards at any normal-high resolution/high quality settings. It's possible there was a bottleneck somewhere in a CF system that held it back, but I just don't think that it would be a CPU bottleneck. Infact I'm willing to bet my Phenom would put out very similar numbers as an overclocked C2D system would with a given videocard on a 24" monitor.

The easiest way to tell is if you see 4850 CF scaling well, there's numerous examples of it scaling close to 100%. When you see 4870 CF scaling much worst and hitting the same FPS as 4850CF then you know you're CPU bottlenecked. There's plenty of examples of this all over and is probably the reason 4870CF hasn't gotten much attention.

Here's a good example.

You can see all the CF/SLI solutions don't scale that well compared to single solutions and all cap out around the same FPS, even across resolutions.

 

[Rebel44]

Originally posted by: chizow

I don't really buy #2... I didn't read everything you and BFG wrote back and forth, so I didn't see if you had any examples of CPU bottlenecking. I could be wrong, and I'd be interested to see some numbers, but I just don't think a modern CPU will bottle neck video cards at any normal-high resolution/high quality settings. It's possible there was a bottleneck somewhere in a CF system that held it back, but I just don't think that it would be a CPU bottleneck. Infact I'm willing to bet my Phenom would put out very similar numbers as an overclocked C2D system would with a given videocard on a 24" monitor.

The easiest way to tell is if you see 4850 CF scaling well, there's numerous examples of it scaling close to 100%. When you see 4870 CF scaling much worst and hitting the same FPS as 4850CF then you know you're CPU bottlenecked. There's plenty of examples of this all over and is probably the reason 4870CF hasn't gotten much attention.

Here's a good example.

You can see all the CF/SLI solutions don't scale that well compared to single solutions and all cap out around the same FPS, even across resolutions.

When I already have more then 60 FPS I dont care if I have 75 instaed of 79....

Also - I can just use higher quality efects and higher AA + AF and still have very good FPS

 

[SlowSpyder]

Originally posted by: chizow

I don't really buy #2... I didn't read everything you and BFG wrote back and forth, so I didn't see if you had any examples of CPU bottlenecking. I could be wrong, and I'd be interested to see some numbers, but I just don't think a modern CPU will bottle neck video cards at any normal-high resolution/high quality settings. It's possible there was a bottleneck somewhere in a CF system that held it back, but I just don't think that it would be a CPU bottleneck. Infact I'm willing to bet my Phenom would put out very similar numbers as an overclocked C2D system would with a given videocard on a 24" monitor.

The easiest way to tell is if you see 4850 CF scaling well, there's numerous examples of it scaling close to 100%. When you see 4870 CF scaling much worst and hitting the same FPS as 4850CF then you know you're CPU bottlenecked. There's plenty of examples of this all over and is probably the reason 4870CF hasn't gotten much attention.

Here's a good example.

You can see all the CF/SLI solutions don't scale that well compared to single solutions and all cap out around the same FPS, even across resolutions.

That's how you determined it to be a CPU bottleneck? Let's look at the 2560 numbers in the link you provided. A single 4850 scores about 74% the FPS that the single 4870 scored. Crossfired 4850's scored about 80% what the CF'd 4870's scored. That's pretty inline, right? Doesn't seem to be any real bottleneck at that res. If you look at the lower res it's a lot closer, but the CF'd 4870's still are faster then the CF'd 4850's by about 4.5%. Something isn't allowing it to scale as well at lower resolution, but I thought that was always the case with SLI/CF. The higher the res, the better results you're more likely to see. Anyway, all of the above is really a moot point as the way you can absolutely tell that the CPU isn't the bottleneck is the fact that all over the graph, at the lower resolution there are framerates well above 89.61 (the CF'd 4870 framerate). The GTX280 scores 99.65FPS. The 3870x2 CF'd (four GPUs total) scores 97.42FPS. How can the CPU be the bottle neck with the CF'd 4870's, holding at 89.xx FPS, yet the same CPU somehow comes up with another 10FPS for the GTX280?

You are right, something is bottlenecking the CF'd 4870's a bit, but I don't see how it is the CPU when all over the graph that same CPU can go well past the 4870CF FPS. Driver problems? PCIE lane bandwidth? Something? I don't know what, but I don't see how it is the CPU.

 

[Janooo]

Originally posted by: chizow

Originally posted by: thilan29
the chip they use is different and so won't necessarily be x8/x8. And in all likelihood, if there was a bottleneck in that part of the card from the 3870x2, ATI would have replaced it this time around.

What do you think the bridge chip does? Unless the PLX gen 2 chip is able to make an x16 PCIE 2.0 slot into an x32 PCIE 2.0 slot then it will result in an x8/x8 PCIE 2.0 on a single slot.

The new inter GPU connection could eliminate any related issues. The card could present itself to the system as a single GPU card.

 

[magreen]

That article and "analysis" by John Peddie was nothing but buying into AMD marketing. The 256-bit memory bus w/ faster ram? A stroke of genius? Gimme a break. Every techie knows you can do that. Making smaller chips that work together? They haven't done it yet. They've yet to release a card to compete with this generation's high end. The 4850/70 are very nice value cards. But don't count your X2 chickens before they hatch. Their proprietary bus instead of XFire is the one thing mentioned in that article that could put them ahead. But it's a complete unknown. "AMD bests nvidia with graphics chip strategy"? FUD.

...damn Peddiephiles.

 

[nitromullet]

Originally posted by: chizow

Originally posted by: thilan29
the chip they use is different and so won't necessarily be x8/x8. And in all likelihood, if there was a bottleneck in that part of the card from the 3870x2, ATI would have replaced it this time around.

What do you think the bridge chip does? Unless the PLX gen 2 chip is able to make an x16 PCIE 2.0 slot into an x32 PCIE 2.0 slot then it will result in an x8/x8 PCIE 2.0 on a single slot.

It really depends on what the PCIe bridge does... It may not provide communication between the PCIe slot and both gpus. It's possible that only one gpu has communication with the PCIe slot and that the bridge chip provides communication between the two gpus. In this type of arrangement, you would essentially have dual sets of PCIe X16 lanes. One set of X16 lanes on on the motherboard, and another set internally on the card.

From looking at the 16x/16x/4x/4x 4-way crossfire results:

http://forums.anandtech.com/me...=2201032&enterthread=y
http://www.tweaktown.com/artic..._crossfirex/index.html

I think it becomes pretty clear that for multi-gpu scaling beyond 2 or 3 gpus, that motherboards are not going to provide enough usable bandwidth to facilitate the communication between all these gpus. Hopefully, ATI has recognized that and is working on a remedy with the 4870X2.

Originally posted by: magreen
That article and "analysis" by John Peddie was nothing but buying into AMD marketing. The 256-bit memory bus w/ faster ram? A stroke of genius? Gimme a break. Every techie knows you can do that. Making smaller chips that work together? They haven't done it yet. They've yet to release a card to compete with this generation's high end. The 4850/70 are very nice value cards. But don't count your X2 chickens before they hatch. Their proprietary bus instead of XFire is the one thing mentioned in that article that could put them ahead. But it's a complete unknown. "AMD bests nvidia with graphics chip strategy"? FUD.

...damn Peddiephiles.

That is a very valid point. This article is a tad bit premature in crowning winners. IMO, if 4870 X2 is to GTX 280 as 3870 X2 was to 8800GTX/Ultra, I wouldn't consider it a clear victory for ATI. That being said, the individual 4870 looks to be a much better card than the 3870 and it beats the GTX 260 in almost every benchmark I've seen, so it does look promising for ATI.

 

[ronnn]

Both sides can win the strategic battle. All this hype is good, now for some games using various new features.

 

[thilanliyan]

Originally posted by: chizow

Originally posted by: thilan29
the chip they use is different and so won't necessarily be x8/x8. And in all likelihood, if there was a bottleneck in that part of the card from the 3870x2, ATI would have replaced it this time around.

What do you think the bridge chip does? Unless the PLX gen 2 chip is able to make an x16 PCIE 2.0 slot into an x32 PCIE 2.0 slot then it will result in an x8/x8 PCIE 2.0 on a single slot.

My fault. You're right there. If in fact the 2 GPUs communicate separately with the bus then it would basically be a x8/x8 configuration with the motherboard. According to the AT article, there's some "Crossfire sideport" that will be enabled for the X2 so hopefully GPU-GPU communication is done that way.

 

[chizow]

Originally posted by: nitromullet
It really depends on what the PCIe bridge does... It may not provide communication between the PCIe slot and both gpus. It's possible that only one gpu has communication with the PCIe slot and that the bridge chip provides communication between the two gpus. In this type of arrangement, you would essentially have dual sets of PCIe X16 lanes. One set of X16 lanes on on the motherboard, and another set internally on the card.

From looking at the 16x/16x/4x/4x 4-way crossfire results:

http://forums.anandtech.com/me...=2201032&enterthread=y
http://www.tweaktown.com/artic..._crossfirex/index.html

I think it becomes pretty clear that for multi-gpu scaling beyond 2 or 3 gpus, that motherboards are not going to provide enough usable bandwidth to facilitate the communication between all these gpus. Hopefully, ATI has recognized that and is working on a remedy with the 4870X2.

It is possible they use a master/slave configuration, but I don't know if its possible. That'd put incredible stress on the master GPU with 2x as many I/Os at least. Still, that wouldn't really solve the bandwidth problem as you'd still have up to 2x as much data passing over a x16 bus compared to 2 GPU on their own x16 slots.

It will be interesting to see what they do with the CF sideport but still, these are clearly things someone should be looking at if interested in the product. I would say for sure that if you have a PCIE 1.1 x16 slot to upgrade to 2.0 before considering this card over a single-gpu solution.

 

[chizow]

Originally posted by: SlowSpyder
You are right, something is bottlenecking the CF'd 4870's a bit, but I don't see how it is the CPU when all over the graph that same CPU can go well past the 4870CF FPS. Driver problems? PCIE lane bandwidth? Something? I don't know what, but I don't see how it is the CPU.

Well I wouldn't focus so much on 5-10FPS differences, especially when comparing to different vendors. Sure there's going to be scaling/driver/vendor/chipset/bus issues that are going to prevent exact results but its much more clear when you compare products from the same vendor. That example and a few others in that review are classic examples of CPU bottlenecking and show you may want to upgrade that CPU before you add that 2nd GPU.

 

[SlowSpyder]

Originally posted by: chizow

Originally posted by: SlowSpyder
You are right, something is bottlenecking the CF'd 4870's a bit, but I don't see how it is the CPU when all over the graph that same CPU can go well past the 4870CF FPS. Driver problems? PCIE lane bandwidth? Something? I don't know what, but I don't see how it is the CPU.

Well I wouldn't focus so much on 5-10FPS differences, especially when comparing to different vendors. Sure there's going to be scaling/driver/vendor/chipset/bus issues that are going to prevent exact results but its much more clear when you compare products from the same vendor. That example and a few others in that review are classic examples of CPU bottlenecking and show you may want to upgrade that CPU before you add that 2nd GPU.

The way I see it, if the CPU was truely the bottleneck then it wouldn't matter what vendor GPU you are using...

One system has x videocard(s) and scores 90FPS. Another system has y videocard(s) and scores 100FPS. Both systems use the exact same CPU. The CPU cannot be bottlenecking the tested game @ 90FPS as it scores well enough above that in the second system. That's how I see it anyway. I guess we'll just have to agree to disagree. I thought 1-2% is within the margin for error, but >10% would not be. I'm hardly an expert on all of this, but it just doesn't seem like a pure CPU bottleneck to me. <shrug>

 

[Tempered81]

I bet R700 > GTX 280 in > 50% of gaming benchies.

 

[nitromullet]

Originally posted by: chizow

Originally posted by: nitromullet
It really depends on what the PCIe bridge does... It may not provide communication between the PCIe slot and both gpus. It's possible that only one gpu has communication with the PCIe slot and that the bridge chip provides communication between the two gpus. In this type of arrangement, you would essentially have dual sets of PCIe X16 lanes. One set of X16 lanes on on the motherboard, and another set internally on the card.

From looking at the 16x/16x/4x/4x 4-way crossfire results:

http://forums.anandtech.com/me...=2201032&enterthread=y
http://www.tweaktown.com/artic..._crossfirex/index.html

I think it becomes pretty clear that for multi-gpu scaling beyond 2 or 3 gpus, that motherboards are not going to provide enough usable bandwidth to facilitate the communication between all these gpus. Hopefully, ATI has recognized that and is working on a remedy with the 4870X2.

It is possible they use a master/slave configuration, but I don't know if its possible. That'd put incredible stress on the master GPU with 2x as many I/Os at least. Still, that wouldn't really solve the bandwidth problem as you'd still have up to 2x as much data passing over a x16 bus compared to 2 GPU on their own x16 slots.

It will be interesting to see what they do with the CF sideport but still, these are clearly things someone should be looking at if interested in the product. I would say for sure that if you have a PCIE 1.1 x16 slot to upgrade to 2.0 before considering this card over a single-gpu solution.

The fact of the matter is that we really don't know exactly what the cause of the slowdown is in 8x/8x Crossfire situations nor do we know how ATI is planning on addressing that on the 4870X2. The way I see it, if a single 4870 X2 in a PCIe x16 slot matches or beats dual 4870s in 16x/16x slots than I really don't care what bridge they use or how they use it.

Your warning/recommendation is pure speculation at this point...

 

[bunnyfubbles]

Originally posted by: chizow

Originally posted by: thilan29
the chip they use is different and so won't necessarily be x8/x8. And in all likelihood, if there was a bottleneck in that part of the card from the 3870x2, ATI would have replaced it this time around.

What do you think the bridge chip does? Unless the PLX gen 2 chip is able to make an x16 PCIE 2.0 slot into an x32 PCIE 2.0 slot then it will result in an x8/x8 PCIE 2.0 on a single slot.

Doesn't PCI-e 2.0 double the bandwidth per lane (500MB/s vs. 250MB/s) over PCI-e 1.x (ie an x16 slot on PCI-e 2.0 is 8GB/s vs. 4GB/s of PCI-e 1.x...)?

I'll admit I haven't kept up with multi GPU technology because I've never really considered it as a viable solution for my own use, but have the multi GPU cards already saturated the single x16 slot of PCI-e 2.0 for this to even be a problem on PCI-e 2.0 motherboards?

 

[BFG10K]

ATI needs to prove this first like RV770 did for price performance. So far they haven't shown anything.

One tends to agree. It's one thing to slap two GPUs onto a single board that appears as one device in Windows Task Manager, but it's another to demonstrate good scaling across a range of apps, not just cherry-picked ones reviewers test.

I didn't read everything you and BFG wrote back and forth, so I didn't see if you had any examples of CPU bottlenecking.

FYI - I ran some tests on my 8800 Ultra a while back. At the settings I used in the games I basically got no gain from going to a E6600 to a E6850. When I dropped back to 1600x1200x4 I got a bit of a gain but only in older CPU limited titles like HL2 and Painkiller.

Here?s the link:

http://episteme.arstechnica.co...007078831#492007078831

Something isn't allowing it to scale as well at lower resolution, but I thought that was always the case with SLI/CF

I haven?t looked the results but SLI/CF are almost always slower in CPU limited situations (i.e. low resolutions) than single cards due to the overhead of multi-GPU synchronization.

When I already have more then 60 FPS I dont care if I have 75 instaed of 79....

So are all cards are equal as long as they?re over 60 FPS?