SLI & CFX PCIe Bandwidth Perf. - x16/x16 vs. x16/x8 @ [H]

I always figured that cards weren't even close to utilizing all of the x16 bandwidth, or even 8 lanes of PCI-E. I wonder if you would have hit the bottleneck if you brought the resolutions and IQ down enough to make the CPU the bottleneck, if you would have seen a difference in performance (I know this is irrelevant to real world game play/what you are all about, but it would be interesting to see).

Having seen this, I would space my 2nd 285 down to my 8x slot if I had a longer SLI bridge to do it with :(.
 
When you do the x16/x16 vs x8/x8 test, I would really love to see how an older mobo does with native x8/x8 like a p45 board with a 775 quad running at about the same clock speed as your i7 920, even if just a few benchmarks for comparison.

I know this is an unfair test, but it would be great to see how much difference an older system has compared a newer one for gaming. We all know that an i7 quad with x16/x16 will be better, but is it signifcantly better? Is it worth upgrading if you can run at x8/x8 on an older system? My understanding is the CPU has the least amount of affect for gaming performance once you reach a certain level of clockspeed (3.4-3.6ish?) but I have never seen any hard numbers to back this up.
 
Very nice overview.:)

Already pretty much determined the conclusion though, as we still don't have cards out that can take full advantage of the extra bandwidth.

I'm running 2x GTX 470's in SLI and it's running x16/x8 and this allows me to spread the cards apart for better cooling.(and in the end as you can see, not take much to any of a performance hit.)
 
bizzmeister said:
Im soo happy Kyle and the crew did this. Im opening my case tomorrow and spacing out my sli gtx 480 sc + setup. WOW. This is great news.
Click to expand...

Yes, this reason alone is why this type of review has been sorely needed. Thanks [H] guys.
 
Good review. Love the 3-monitor resolutions. It's all I look for anymore and it's valuable to me when they're there.
 
GoodBoy said:
Yes, this reason alone is why this type of review has been sorely needed. Thanks [H] guys.
Click to expand...

Building these GTX 480 SLI Fermi systems lately what was got me really thinking about this being "needed" about now.
 
Last edited:
In reference to the GTX 460 SLI versus RadeonHD 5870 Crossfire performance:

RadeonHD 2900: May 2007: http://www.hardocp.com/article/2007/05/13/ati_radeon_hd_2900_xt/
RadeonHD 3870 / 3850: November 2007 :: http://www.hardocp.com/article/2007/11/15/ati_radeon_hd_3800_series

Time Difference: 5 months
Performance Difference: Huge

RadeonHD 3870: November 2007
RadeonHD 4870 / 4850: June 2008 :: http://www.hardocp.com/article/2008/06/25/amd_radeon_hd_4800_series/

Time Difference: 8 months
Performance Difference: Huge

We are very excited to see a brand new video card in the $199 domain that simply smokes the previous generation in terms of the gameplay advantage provided. Not only that, but it is an AMD ATI GPU based video card doing it
Click to expand...

I think some people can already see where this is going.

RadeonHD 4870: June 2008
RadeonHD 5870: September 2009: http://www.hardocp.com/article/2009/09/22/amds_ati_radeon_hd_5870_video_card_review/

Time Difference: 13 months
Performance Difference: Huge

RadeonHD 5870: September 2009
Geforce GTX 460: July 2010
Time Difference: 10 months

Performance difference; Single GPU, RadeonHD is more powerful, multi-gpu, GTX 460 is more powerful.

Here's the thing: The RadeonHD 5800 series is just one month away now from being on the market for a full complete year. Ignoring the RadeonHD 2x00 series and the GeforceFX series for a minute, most of the time when a vendor releases a mid-range card that's 7 months or more out from a high-end card, that mid-range card is able to stomp the previous high-end card into the ground.

With 10 months separation between the products? Seriously, AMD probably already had plans to just drop the 5830, put the 5850 at the $200 mark, and drop the price out of the 5870. They wound up not needing to really move their products at all. Nvidia's respin turned out to be a worthy competitor... to a filler card. Let me repeat that: in single GPU rendering, which is what most people buy, the first respin on Fermi cannot reliably outperform a card that was released as a FILLER. Comparing the GTX 460 to any of the HD 58xx series is a bit of a mugs game. RV870 is not the card's competitor.

Rather, the GTX 460 is the first respin out of Nvidia's development, and it's one they desperately needed. GTX 460 (GF104) does for Fermi (GF100) what RV670 did for R600. The RadeonHD 2x00 series was a complete dog of a release. The RadeonHD 3x00 series was something you'd actually want to spend money on.

Ergo, the GTX 460, as a respin, needs to be compared to the respin of RV870. Nvidia has had time to see what the R800 series could do, time to figure out how to those features better, time to implement those features, and so on... and the end-product still can't actually out-run a card released nearly 5 months before-hand in Single-GPU rendering modes.

A really big problem for Nvidia is that they won't get the chance of a respin on RV870.. AMD basically has already said there won't be a product refresh on RadeonHD 5000 series, they will instead go straight to RadeonHD 6000 series. Nobody, outside of AMD, knows what the next series of GPU's will be like, or what they can do.

We can extrapolate, based on the previous releases, that a RadeonHD 6000 series card will likely double performance compared to previous releases while retaining a similar power-usage envelope. Then again, that's pretty much AMD's entire strategy, GPU and CPU. Increase processing power and IPC within thermal and power limitations.

This is why AMD is been pretty much ambivalent on GTX 460 cards out-running the RadeonHD 5870 cards in multi-gpu modes. AMD was probably expecting the Fermi derived respins to probably dominate RadeonHD 5x00 series in single-card performance as well. When GF104 hit and it turned out to just be a mid-range card that was only superior in multi-gpu modes, AMD's engineers were probably laid out flat laughing. Nvidia took aim, and shot at, a target that was already out-of-date.

Now, okay, fine, if you are buying new, right now the GTX 460 is astounding card. As far as mid-range cards go, it's an excellent performer. It, however, does not do what other mid-range cards do with the amount of time separation between competing products. There's also no getting around that it's competing against an architecture that largely a full-year old...

Right off hand, if I was Nvidia, or if I was an Nvidia supporter, I would be panicking that GF104 doesn't perform better than it does. I would be frightened that my main competitor said they weren't going to bother refreshing their product line-up.

That being said, while things look astoundingly bad for Nvidia right now...

I am a little concerned over AMD's multi-gpu rendering tech. I'm curious as to how much of the performance differences between GTX 460 and HD 5870 in Multi-GPU modes is down to software overhead and load-balancing issues versus an actual hard-ware limitation in how much data can be transferred between the cards and if there's a bottleneck in the RV870's actual chipset design.
 
artika23 said:
Kyle, thank you for adding the 5760 x 1200 benchmarks!
Click to expand...

YES THANK YOU. this has been something that i haven't felt was addressed very much lately. I mean in theory i can take your normal measurements you do and divide by 3 and i'll get roughly the answer to my display setup, but still. Doesn't seem like the triple screen results are published often enough. while i enjoy other sites, canned benches are just that, i would like a real world example, so finally, thank you.
 
This review was ridiculous. You completely ignored Voodoo SLI benchmarks. nothing less than atrocious. Deeply Deeply disappointed in [H]
 
@Saist

Well though out good read.

I still have my suspicions as to the over-all performance of my new Nvidia Asus 460's in SLI. I am really just using logic, some of the reviews I've skimmed over and mostly relying on a big chunk of the data from Hardocp to move away from one of my 5870's.

For me, this is really about maintaining value and trying to keep around the same performance. I just can't afford to lose money on a video card. I have to constantly shift ( sell ) my cards around about every 12 to 14 months.

I am and always will be an early adopter.

I doubt the new 6xxx will beat a pair of 480's in SLI but ... never say never. I'm praying that it does. It's been forever since I've seen a knockout blow from AMD.

I remember back in the day that in the CPU forums here on Hardocp, that the Intel thread was dead .. DEAD for two - three years until the C2D / 6600 came out. Would like to see that happen with AMD again. CPU or GPU, take your pick.

Think I have $850 or so dollars invested into my crossfire 5870's. Selling one nearly paid for a pair of the 460's as already mentioned. The plan is to turn them into the Asus TOP versions as the hardware / cooling solution is supposed to be exactly the same between the two. That will absolutely give me better gaming performance than what I had.

The 2nd card I will sale in a new system I just built. All in all, I will have my cards used, sold and re-invested with little to no investment without missing a beat at the forefront of high-end graphic cards.
 
Ergo :) , we focus on what we think our readers want to know about and what they can BUY today that is a great value. The 5850 held the "best value" title with us for a very long time. The GTX 460 now holds that spot. I guess we will see in October what changes if any we see on the GPU landscape.
 
xjmtx said:
YES THANK YOU. this has been something that i haven't felt was addressed very much lately. I mean in theory i can take your normal measurements you do and divide by 3 and i'll get roughly the answer to my display setup, but still. Doesn't seem like the triple screen results are published often enough. while i enjoy other sites, canned benches are just that, i would like a real world example, so finally, thank you.
Click to expand...

Kyle, this is something that can't be said enough. Thanks for the extra effort to include surround benchmarks. For me, they answered a question I've been asking looking at monitor purchases to with a GTX 460 SLI setup. The answer is yes, 3x 1920x1200 is too big to run modern games at a framerate I'll like. My planned purchase of 3x 1680x1050 is ratified (3x 1920x1080 doesn't make sense to me, I only care about vertical size with a surround setup).
 
Great review!!!

I applaud the fact that you guys used the tape method on a board with only 2 PCIe2.0 x16 slots. It eliminated any variables that might spring up due to the NF200 chip found on boards rated for Tri-SLI.
 
Saist said:
Ergo, the GTX 460, as a respin, needs to be compared to the respin of RV870. Nvidia has had time to see what the R800 series could do, time to figure out how to those features better, time to implement those features, and so on... and the end-product still can't actually out-run a card released nearly 5 months before-hand in Single-GPU rendering modes.
Click to expand...

It gets worse. It is a larger chip than 5870 and quite slower at the same time. It costs Nvidia more money to make a 460 than ATI to make a 5870, but look at the prices. Nvidia is either loosing money on every 460 sold or, at best, making minimal profit (depends on the exact yield from TSMC). Probably the only reason ATI hasn't lowered 5850/5870 prices is that they sell all they can make anyway.

GTX 460 is a great card for us customers, but a terrible one for Nvidia (entire Fermi line unfortunately). Then there is soon to be released Southern Islands from ATI, and then NI next year. Ouch!

As much as I am glad that ATI is doing this great, it is terrible (for us) that Nvidia is this uncompetitive and without any apparent (financially viable) solution in the near future.
 
Though I do appreciate the effort that went into this review I do believe there is one item of contention. That is the CPU speed. A 3.6 Ghz OC for an SLI system would most likely end up being the bottleneck. Essentially all of your results are close together because your CPU isn't clocked high enough to max out even the x8 PCIe lane.

I wouldn't expect a complete revamp of the tests but perhaps a single test OCing the CPU to 4.0 or higher would be enough to show the point at which the x8 PCIe becomes a bottleneck.

It's possible that at 3.6 mark you could see the bottleneck for a x4 PCIe lane but probably not an x8 lane. I think a 4.0 OC is within most OCPer's reach, is there a possibility to get one of the tests redone to incorporate a higher clock on the CPU?
 
3far4shot said:
Already pretty much determined the conclusion though, as we still don't have cards out that can take full advantage of the extra bandwidth.
Click to expand...

I don't like being repetitive, but...there's no need for any card today to "take full advantage of the extra bandwidth" that PCIex16 provides, because the local bandwidth on today's 3d cards (such as those tested for this [H] article) is a multiple of times greater/faster than the bandwidth provided by PCIex16. You notice the slowdowns during those times when a card is forced to access the PCIe bus (at any speed) as opposed to accessing its own local bus. That's why there's no essential difference in performance between x16/x16 and x16/x8. All today's cards are configured to run out of their local buses as much of the time as it is possible for them to do.

The slowest link in the chain is PCIex16, and that's exactly why today's 3d cards come with no less than 512MBs of ram, all the way up to 2GBs. If they had no local ram, or even up to 128MBs to 256MBs, today's cards at decent resolutions would all run dog slow on top of the PCIex16 bus, not to mention an 8xPCIe bus. The local bandwidth internal to a 3d card and the amount of local ram on the card's local bus (no less than 512MBs) is what counts, as the total bandwidth afforded by PCIex16 is but a fraction thereof:

Peak memory bandwidth on an HD 5870 is 153.6 GBs/second. On a GTX 480, peak memory bandwidth is 177.4 GBs/second.

Maximum peak bandwidth for PCIex16 is a concurrent 8GBs/second (bi-directional, the most bandwidth possible for PCIex16 in a single direction is 4GBs/second.)

http://www.intel.com/design/chipsets/pciexpress.htm

So, an HD 5870 has 19.2x the bandwidth of concurrent PCIex16, and a GTX480 has 22.175x the concurrent bandwidth of PCIex16.

Again, 3d cards are designed and engineered, both in terms of their hardware and their driver software, to run as much as possible out of their local buses without having to access the PCIex16 bus, and that's why we have the performance we do from them. Advances in GPU local memory bus bandwidth will always remain a dramatic multiple of the coming advances in system bus technology,imo.

If it is true that PCIe 3.0 will double the peak bandwidth of the current PCIe 2.x spec, then all this will mean is that a current HD 5870 or GTX 480 will still remain ~10-11X faster than PCIe 3.0.
 
WaltC said:
The slowest link in the chain is PCIex16, and that's exactly why today's 3d cards come with no less than 512MBs of ram, all the way up to 2GBs. If they had no local ram, or even up to 128MBs to 256MBs, today's cards at decent resolutions would all run dog slow on top of the PCIex16 bus, not to mention an 8xPCIe bus.
Click to expand...
I find what you said interesting. So when choosing a video card, you would want the most memory (on-board) for future proofing down the road in some sense right? So when deciding a video card, always get the most memory available for that GPU at the time?
 
I would be interested to see the benchmark results if you taped up a PCIe lane one by one to sort out how much bandwidth cards are actually using. ie: Tape up a x16 card to x8, then x7, x6, x5 etc and run a bench at each set of lanes.
 
I have a Gigabyte GTX 460 but I noticed that the card only has 1 SLI connector which means that it can only have a 2 card SLI setup. Do ANY of the GTX 460's have 2 SLI connectors? Will the proposed 2GB GTX 460's have 2? If not then this may be a limitation on usability in the future and a boon to those that have 5870's. Yes I am aware that the third card in crossfire will not scale that well but for those that will be selling their 5870's cheap they may be a good deal for 3 and 4 card crossfire users.....just sayin'.

However if the new GTX 460's have 2 SLI connectors then those will sell very well. I may even sell my Eyefinity 6 cards and then not have to worry about these extra dongles if I want to power 6 monitors. AND the scaling would be sweet.
 
configurations can only be 1 4 8 16x (maybe 2x as well?)

anandtech (or an other site, guessing anandtech would be the only site to do it really) did do the taping thing on the PCI-E card them self's Nvidia card's would work on Boot even when at 1x (ATI cards i think had to be 4x to boot the system) but even at 4x performance was ok but FPS was loss at that point due to bandwidth limits (going to find that now i correct any of the above)

4x/8x is Plenty for even the High end cards (GTX295 or higher you need 8x maybe), you have to understand the idea of an video card is so it does the work not the PC at most the link to the system is giving commands to the GPU at high speed as all the textures are on the card once the game has loaded them up onto the video card, and if your doing SLI or CF the mirroring of the GDDR is done over the SLI or CF bridge not over the PCI-E

LOL it was tomshardware that did it (try not to goto that site any more {if i do not load there site they get no ad support from me} as it went down hill about 3-5 years ago with poor reviews not well thought out reviews or tests)
http://www.tomshardware.com/reviews/sli-coming,927-2.html
checking for others sites
 
Last edited:
I just want to say THANK YOU
I have heard this stated many times but you put the numbers up to show its true.
 
Great review, love the 8x results and the 5760 resolutions.

Can we maybe get 16x/4x for the people with stock P55 boards?
 
I'm currently running 2 GTX480's in SLI and have a PCIEX slot to spare, my MB supports x16 x16 x1 or x16 x8 x8, I have a GTX 260 lying around that want to dedicate to PhysX, which configuration would work best, both GTX 480's in x16 lanes and the GTX260 in x1 or the primary GTX480 in x16 and the secondary GTX480 and the GTX260 in x8?

Based on this review I'm guessing x16 x8 x8
 
Last edited:
Good article thanks ! . I am putting together a new system that will have GTX470s in SLI. I was looking at aftermarket cooling to keep the noise and heat down . Now that I see I will not be losing any performance I will go ahead a space them out right from the start and not worry about it.
 
You must log in or register to reply here.
Back
Top