Crossfire and sli...mobos still really confused... :-(
- Thread starter kim_51
- Start date
wolfram23
Splendid
- Apr 12, 2010
- 6,313
- 0
- 26,460
You want to quadfire 4 cards, or 2 cards with 2 gpus like CF 5970s??
Quadfire even Trifire doesn't scale well. TriSLI scales a little better in theory, but honestly you get very little performanc over a 2 card set up.
That said, the 1366 supports up to 36 PCI lanes. You could get 16x/16x/4x or 16x/8x/8x. The south bridge also has a 4x PCIe 2.0 lane which you can put a GPU into. The 4x lane does still scale performance up but it's only about 80% bandwidth of a 16x lane, while an 8x lane is like 96% of a 16x lane. As you can see, putting anything in a 4x lane is a bit of a waste.
Also, with USB3 and SATA 3 you'll be taking up some of the DMI bandwith which communicates between the PCH (south bridge) and the CPU/IMC so again, a GPU on that PCIe lane could suffer even more.
I haven't researched specific X58 boards, but hopefully you keep these things in mind when choosing.
Quadfire even Trifire doesn't scale well. TriSLI scales a little better in theory, but honestly you get very little performanc over a 2 card set up.
That said, the 1366 supports up to 36 PCI lanes. You could get 16x/16x/4x or 16x/8x/8x. The south bridge also has a 4x PCIe 2.0 lane which you can put a GPU into. The 4x lane does still scale performance up but it's only about 80% bandwidth of a 16x lane, while an 8x lane is like 96% of a 16x lane. As you can see, putting anything in a 4x lane is a bit of a waste.
Also, with USB3 and SATA 3 you'll be taking up some of the DMI bandwith which communicates between the PCH (south bridge) and the CPU/IMC so again, a GPU on that PCIe lane could suffer even more.
I haven't researched specific X58 boards, but hopefully you keep these things in mind when choosing.
bilbat
Splendid
- Sep 21, 2008
- 4,568
- 10
- 23,160
80%? 96% ???
A PCIe lane is 5Gb/s. Period. To figure out GB/s, you, like almost all transmission schemes, lose two bits per frame, so you divide by ten. One lane = ½GB/s. Four lanes = 2GB/s. Eight lanes = 4GB/s. Sixteen lanes = 8GB/s. Period.
No board I am aware of 'splits out' the ICH lanes to anything but x1 slots - they are mainly used up for peripherals: LAN chips, extra SATA controllers,, and, as I mentioned, any x1 slots the board has.
The 58's two sixteens are usually on a 'bifurcation switch', which allows the board to 'negotiate' with its installed cards to do: 2 x 16, 1 x 16 + 2 x 8, or 4 x 8. If you put anything into these slots, even a x1 card, it 'uses up' at least eight lanes - it's how the 'bifurcator' works! Very few MOBOs are physically set up to support four double-width vidcards. Most of those that do, get four sixteens through the use of nVidia NF200 PCIe switch/multiplexors. You cannot put two X58's on a board; there are boards (server/workstation class) that do get seventy-two 'real' full-width, 'raw' lanes (Tyan's 7025, for example) through the use of a pair of 5520 IOH's (the server/Xeon world's 'equivalent' to the X58), but, as a rule, these do not support SLI/xFire - the trick to the 'twinning' is that the 5520, unlike the X58, supports two QPI links. These will also not accept an i7; they require Xeons which also support two physical QPI...
One issue that you need to know, to decide what to spend, for 'how much board', is the vidcards themselves that you intend to use. Very, very few vidcards can actually 'use up' a sixteen lane slot - it's a hell of a lot of bandwidth! In the ATI world, it's pretty much limited to the 5850 (barely...), 5870 (especially the 2Gig cards), and 5970. For nVidias, I'm not as clear, as I'd not fiddle with one, but I'm pretty sure the list starts at the GTX-285, and includes the GTX 470's and 480's... Another thing to be aware of, is that if you wish to use four double-width cards, without water-cooling them, you typically need a fussier case - at least eight slots, and, for many, like the GB UD9, a nine-slot...
A PCIe lane is 5Gb/s. Period. To figure out GB/s, you, like almost all transmission schemes, lose two bits per frame, so you divide by ten. One lane = ½GB/s. Four lanes = 2GB/s. Eight lanes = 4GB/s. Sixteen lanes = 8GB/s. Period.
No board I am aware of 'splits out' the ICH lanes to anything but x1 slots - they are mainly used up for peripherals: LAN chips, extra SATA controllers,, and, as I mentioned, any x1 slots the board has.
The 58's two sixteens are usually on a 'bifurcation switch', which allows the board to 'negotiate' with its installed cards to do: 2 x 16, 1 x 16 + 2 x 8, or 4 x 8. If you put anything into these slots, even a x1 card, it 'uses up' at least eight lanes - it's how the 'bifurcator' works! Very few MOBOs are physically set up to support four double-width vidcards. Most of those that do, get four sixteens through the use of nVidia NF200 PCIe switch/multiplexors. You cannot put two X58's on a board; there are boards (server/workstation class) that do get seventy-two 'real' full-width, 'raw' lanes (Tyan's 7025, for example) through the use of a pair of 5520 IOH's (the server/Xeon world's 'equivalent' to the X58), but, as a rule, these do not support SLI/xFire - the trick to the 'twinning' is that the 5520, unlike the X58, supports two QPI links. These will also not accept an i7; they require Xeons which also support two physical QPI...
One issue that you need to know, to decide what to spend, for 'how much board', is the vidcards themselves that you intend to use. Very, very few vidcards can actually 'use up' a sixteen lane slot - it's a hell of a lot of bandwidth! In the ATI world, it's pretty much limited to the 5850 (barely...), 5870 (especially the 2Gig cards), and 5970. For nVidias, I'm not as clear, as I'd not fiddle with one, but I'm pretty sure the list starts at the GTX-285, and includes the GTX 470's and 480's... Another thing to be aware of, is that if you wish to use four double-width cards, without water-cooling them, you typically need a fussier case - at least eight slots, and, for many, like the GB UD9, a nine-slot...
wolfram23
Splendid
- Apr 12, 2010
- 6,313
- 0
- 26,460
I'm sure you'll find this interesting, as this is where my info on the 16x vs 8x vs 4x comes from:
http://www.tomshardware.com/reviews/p55-pci-express-scaling,2517-3.html
Here's a quick snapshot of some of the results:
http://www.tomshardware.com/reviews/p55-pci-express-scaling,2517-3.html
Here's a quick snapshot of some of the results:
bilbat
Splendid
- Sep 21, 2008
- 4,568
- 10
- 23,160
You are confusing bandwidth with framerate, and or 'synthetic' benchmarking. While the bandwidth has some effect on framerate (mainly on vidcards that are fast enough to use high bandwidth), framerate is a product of a huge number of other variables: GPU architecture and relative rendering efficiency; structure and utility of the PCIe 'bridge' chip(set) connecting to the PCIe; VRAM speed and bus width; the speed at which the vidcard can DMA to the 'mapped' main memory; utility of the drivers; the optimizations made in the 'game engine' for the driver/GPU architecture combination; and on and on... Even while holding as many of these as contstant as can be managed, all those results prove is exactly what I said - very little on the market can take any advantage of a x16 slot, and the few that can, just barely push the limits of x8... Think of it this way - if you have a 3/4" faucet on the side of your house, you will get a little more water by using a 3/4" hose, than by using a 5/8" hose - but, there will be no increase in flow by going to a 1" hose, or a 2" hose - 'cause you just can't supply any more water!
If your video card can only produce a 'stream' that's 4½ GB/s, it will be slightly 'limited' by a 4GB/s x8 slot, but doubling that to an 8GB/s x16 slot is never going to get you more than the 4½ GB/s the the card can produce! It doesn't mean that the actual bandwidth of the x16 is anything but double that of the x8 - it simply means that the 'extra' bandwidth can't all be used... This is the situation with most of the hardware busses in current use in our PCs - people are seduced by that 'bigger number' at the end of SATA3, but it's, for the main part, useless. No physical HD can come close to the bandwidth limits of SATA2, even SSD's can't on any kind of a sustained basis (but there is hope, with the new 'sandforce' controllers!); and the driver/architectural optimizations made by half a decade of development of the ICH's make the Marvell controllers pale by comparison, but the poor sheep paying doesn't understand bandwidth from bupkiss - all he 'sees' is the bigger number at the end - and that's all the marketing guys want him to see!
If your video card can only produce a 'stream' that's 4½ GB/s, it will be slightly 'limited' by a 4GB/s x8 slot, but doubling that to an 8GB/s x16 slot is never going to get you more than the 4½ GB/s the the card can produce! It doesn't mean that the actual bandwidth of the x16 is anything but double that of the x8 - it simply means that the 'extra' bandwidth can't all be used... This is the situation with most of the hardware busses in current use in our PCs - people are seduced by that 'bigger number' at the end of SATA3, but it's, for the main part, useless. No physical HD can come close to the bandwidth limits of SATA2, even SSD's can't on any kind of a sustained basis (but there is hope, with the new 'sandforce' controllers!); and the driver/architectural optimizations made by half a decade of development of the ICH's make the Marvell controllers pale by comparison, but the poor sheep paying doesn't understand bandwidth from bupkiss - all he 'sees' is the bigger number at the end - and that's all the marketing guys want him to see!
thanks dude..... i'll keep these things in mind while getting a mobo... but problem is that i always go for future... thats why i want to get latest features...
and please tell me about
asus rampage 3 mobo.....is it good and full of latest features?
wolfram23
Splendid
- Apr 12, 2010
- 6,313
- 0
- 26,460
bilbat: I'm sure you're right as that does completely explain why 8x was only 4% worse. However, it still remains that a 5870 is a top end card and loses 4% of it's speed on an 8x lane. While your theory is totally right, in practice that is the result. Now, I can honestly say I never really considered it the way you explain it, so I suppose at a certain point in the GPU heirarchy, the cards won't see any difference on a 8x lane. Is it GTX460 and lower? Maybe GTS275 or 5770. I don't know but it's interesting to think about.
On the other hand, wouldn't a 5970 get severely limited by an 8x lane?
Regardless, you've raised some good points. Sata3 and USB3 aren't entirely useful at this point but maybe in a year we'll have some super high bandwidth devices.
On the other hand, wouldn't a 5970 get severely limited by an 8x lane?
Regardless, you've raised some good points. Sata3 and USB3 aren't entirely useful at this point but maybe in a year we'll have some super high bandwidth devices.
TRENDING THREADS
-
News US sanctions transform China into legacy chip production juggernaut — production jumped 40% in Q1 2024- Started by Admin
- Replies: 34
-
-
Question New pc build r9 7900x3d rtx 4080 super no post only ram rgb turns on- Started by Harvey Durward
- Replies: 4
-
RTX 4070 vs RX 7900 GRE faceoff: Which mainstream graphics card is better?
- Started by Admin
- Replies: 72
-
-
News TSMC to charge premium for making chips outside of Taiwan, including its new US fabs, CEO says
- Started by Admin
- Replies: 22
Facebook
Twitter
Instagram