16x PCI-e vs 8x PCI-e in CrossFire/SLI (The Answer..)

[zerosaku]

It seems many threads all over the internet has/had/have asked this same question all over.

was tweaktown's review really truthful ? sadly.. it wasn't.

Difference between 16x PCI-e vs 8x PCI-e in CrossFire/SLI ?

Answer:

Here is why.

PCI-e 2.0 16X = 8GB/s total bandwidth
PCI-e 2.0 8X = 4GB/s total bandwidth

ATI:

HD 3870 = 1125mhz (dual piped - 2.2 GB/s data rate) = 2.2 GB/s total data rate
HD 4650 = 500mhz (dual piped - 1 GB/s data rate) = 1 GB/s total data rate
HD 4670 = 1000mhz (single piped - 2 GB/s data rate) = 2 GB/s total data rate
HD 4770 = 800mhz (quad piped - 3.6 GB/s data rate) = 3.6 GB/s total data rate
HD 4830 = 900mhz = (dual piped - 1.9 GB/s data rate) = 1.9 GB/s total data rate
HD 4850 = 993mhz (dual piped - 1.9 GB/s data rate) = 1.9 GB/s total data rate
HD 4850 X2 = 993mhz (dual piped - 1.9 GB/s data rate)(times 2) = 3.8 GB/s total data rate
HD 4870 = 900mhz (quad piped - 3.6 GB/s data rate) = 3.6 GB/s total data rate
HD 4890 = 975mhz (quad piped - 3.9 GB/s data rate) = 3.9 GB/s total data rate
HD 4870 X2 = 900mhz (quad piped - 3.6 GB/s data rate)(times 2) = 7.2 GB/s total data rate

So as you can see, the 4870 X2 will be hurt in an 8X condition and why the 3870, 4650, 4670, 4770, 4830, 4850, 4870, 4890, and 4850 X2 will not be hurt in performance at all.

Nvidia:

8800 Ultra = 1080mhz (dual piped - 2.1 GB/s data rate) = 2.1 GB/s data rate
9600 GT = 900mhz (dual piped - 1.8 GB/s data rate) = 1.8 GB/s total data rate
9800 GTX = 1100mhz (dual piped - 2.2 GB/s data rate) = 2.2 GB/s total data rate
GTS 250 = 1100mhz (dual piped - 2.2 GB/s data rate) = 2.2 GB/s total data rate
GTX 260 216 = 999mhz (dual piped - 1.9 GB/s data rate) = 1.9 GB/s total data rate
GTX 275 = 1134mhz (dual piped - 2.2 GB/s data rate) = 2.2 GB/s total data rate
GTX 280 = 1053mhz (dual piped - 2.1 GB/s data rate) = 2.1 GB/s data rate
GTX 285 = 1242mhz (dual piped - 2.4 GB/s data rate) = 2.4 GB/s data rate
GTX 295 = 999mhz (dual piped - 1.9 GB/s data rate)(times 2) = 3.8 GB/s total data rate
GTX 285X2 = 1242mhz (dual piped - 2.4 GB/s data rate)(times 2) = 4.8 GB/s data rate

So as you can see, the GTX285 X2 will be hurt in an 8X condition and why the 8800, 9600, 9800, 250, 260, 280, 275, 285 & 295 will not be hurt in performance at all.

Check this pic:


This should be a clear message to anyone wanting to do multi-gpu setup's, people with PCI-e 2.0 @ 16x-16x-8x or 16x-8x-8x or even 8x-8x-8x bandwidth should have no bottle neck for any single gpu cards in the market, aside from the double gpu cards like 4870x2, and incoming GTX285x2.

a quick example:

Which is better performance in fps/gaming, 16x-8x-8x w/ PCI-e 2.0 motherboard vs 16x-16x-16x w/ PCI-e 2.0 motherboard vs 8x-8x-8x w/ PCI-e 2.0 motherboard:

1x 4870 1GB @ 16x slot
1x 4870 1GB @ 8x slot
1x 4870 1GB @ 8x slot
vs
1x 4870 1GB @ 16x slot
1x 4870 1GB @ 16x slot
1x 4870 1GB @ 16x slot
vs
1x 4870 1GB @ 8x slot
1x 4870 1GB @ 8x slot
1x 4870 1GB @ 8x slot

Winner:
Draw, no difference in fps. Although some 16x-16x-16x boards with n200 chipset's has fps drops in games which is reviewed many times but no drops in benchmarking (3dmarks,vantage). If your board is PCI-e 1.0 then that would be a huge difference :)

[C3D76]

Re: 16x PCI-e vs 8x PCI-e in CrossFire/SLI (The Answer..)

Good info Zero.

Nice and easy to see all the numbers like this :).

Just noticing 4870X2=7.2GB/s, so 4890X2 would be 7.8GB/s .?.
With PCI-e 2 x16 @ 8GB/s, we're gonna need more bandwidth (PCI-e 3) very soon >.<.

GTX 285X2 o.0!
Only thing I found was the Asus Mars. Is that what you meant?

[zerosaku]

Re: 16x PCI-e vs 8x PCI-e in CrossFire/SLI (The Answer..)

Yep, but there is no 4890x2 in the market...

and yes gtx 285x2 = Asus Mars

[randomness]

Re: 16x PCI-e vs 8x PCI-e in CrossFire/SLI (The Answer..)

theres an article over at tomshardware comparing PCI-e x16/8/4/2/1 crossfire speeds and single card speeds.

go check it out

[largon]

Re: 16x PCI-e vs 8x PCI-e in CrossFire/SLI (The Answer..)

It seems many threads all over the internet has/had/have asked this same question all over.

was tweaktown's review really truthful ? sadly.. it wasn't.

EXCEPT that your numbers make no sense whatsoever.

You have listed the memory frequency of many video cards and somehow linked the bandwidth-per-pin, that is, the bandwidth per 1bit of memory bus of the cards' VRAM to PCIe link bandwidth and made deductions on bottlenecking based on this invalid comparison.

FACT IS, the video card -specific bandwidth numbers listed in OP have absolutely nothing to do with PCIe bandwidth. There's no sense at all to compare them and no conclusions whatsoever can be made based on them.

=)