860 v 920 - BUS speed!

[hioahaa]

So everyones been raving about the 860 being so much better than the 960. but is it really????


no one in their arguments seems to point out the difference in bus speed - im not a haXx0r so i'll quote something from villagegeek.com...

'QPI provides a wider PCIe bus at 20 lanes bidirectional (40 single lanes), where DMI has more standard 16 lane PCIe bus. Since the entire bus is built on PCIe standards now, the difference is significant.

When all the calculating is done, the QPI interface can transfer 25.6 GB/s which is exactly twice the rate of the fastest previous hub based motherboard. The DMI interface claims to handle 10GB/s although I was unable to confirm that specification on Intel?s website.

So the i7 is 2.66GHz clock speed, 4.8 GT/s per second, and uses the QPI interface which transfers up to 25.6 GB/s of data.

While the i5 has the same 2.66 GHz clock speed, it transfers 2.5 GT/s and the DMI interface can handle less than half the data across the motherboard.'

http://www.villagegeek.com/Arc...d_i5_architecture.html

So....if the 920 can can handle double the data across the mobo, surely in practice that makes it a lot faster than the 860 even though its clock speed is lower????

can someone please shed some technical know how and light on this, and explain it in practical laymens terms!
is bus speed a major limiting factor?

this explains it pretty well:
http://www.hardwaresecrets.com/printpage/708

 

[masteryoda34]

You are correct. However, if nothing actually uses more than 10GB/s, then you won't be able to tell the difference. Since in most testing you can't tell the difference, we know that 25.6GB/s of bus bandwidth for the i7 is really more than is necessary. QPI was made for dual and quad socket server systems. Desktop's dont really need that much bandwidth.

 

[Nemeth782]

That article has some significant flaws...

Firstly, QPI in a non-extreme edition i7 runs at 4.8GT/s and yields a data rate of 19.2 GB/s. The 25.6GB/s figure is for the 6.4GT/s QPI of the 965 and 975.

This 19.2GB/s is used to communicate with the x58 chipset, and all data goes via the chipset. All graphics cards are attached to PCIe lanes on the x58, as are SATA controllers, gigabit ethernet ports, USB, etc etc.

Then you look at Core i5 and i7 8xx.

They do not use QPI, instead using DMI which is basically PCIe x4 v1.1, to communicate with the P55 chipset. This gives them 10GB/s bandwidth to the chipset. Interestingly DMI was previously used to link older Core 2 Northbridges to the Southbridge, like P45 to ICH10R etc.

However, the i5 and i7 8xx do not communicate with everything via the P55 chipset, crucially the majority of the PCIe lanes are actually on the CPU - the graphics card(s) are connected directly to the CPU.

The P55 just does USB, SATA, and a couple of the PCIe x1 slots.

As such, there is alot less need for bandwidth between P55 and i5/i7 8xx than there is between x58 and i7 9xx

 

[Denithor]

Yoda hit it right on the head - i7 was designed with more bandwidth than it needs for "desktop" applications so you won't notice any difference at all in 99.9% of situations you are likely to encounter in a typical home use environment.

Now, if you're building a server level machine the increased bandwidth - of both the QPI and triple channel memory - will be crucial to higher performance. But otherwise just grab an i5 750 and be happy.

 

[lopri]

QPI isn't necessary on P55 since Lynnfield has on-die PCIe controller. (though it's there for 'licensing' purposes) Both Bloomfield and Lynnfield CPUs talk directly to memory so that leaves video, networking, storage, etc. are where QPI/DMI come into play on single-socket systems. I don't like that Intel's DMI has been stuck for generations (probably due to compatibility reasons on OEM designs), but in Intel's defense your average SATAII drives are not bottlenecked by DMI interface. DMI is also used on X58 system to connect north bridge and south bridge, so the comparison of GB/s becomes even more moot. Other than video card slots and maybe one another PCIe slot that's rooted in X58 north bridge, everything else is attached to ICH10R which transfers data via DMI to the north bridge (then to the CPU).

 

[IntelUser2000]

Originally posted by: Nemeth782
They do not use QPI, instead using DMI which is basically PCIe x4 v1.1, to communicate with the P55 chipset. This gives them 10GB/s bandwidth to the chipset. Interestingly DMI was previously used to link older Core 2 Northbridges to the Southbridge, like P45 to ICH10R etc.

DMI on P55 is 2GB/s(http://www.intel.com/Assets/PDF/prodbrief/322641.pdf). Still, that won't bring performance differences as applications in desktops aren't limited by the bandwidth of the hard drive and peripheral I/O.

 

[Nemeth782]

Originally posted by: IntelUser2000

Originally posted by: Nemeth782
They do not use QPI, instead using DMI which is basically PCIe x4 v1.1, to communicate with the P55 chipset. This gives them 10GB/s bandwidth to the chipset. Interestingly DMI was previously used to link older Core 2 Northbridges to the Southbridge, like P45 to ICH10R etc.

DMI on P55 is 2GB/s(http://www.intel.com/Assets/PDF/prodbrief/322641.pdf). Still, that won't bring performance differences as applications in desktops aren't limited by the bandwidth of the hard drive and peripheral I/O.

You are right, as it is basically PCIe 1.1 x4, so 1Gbyte/sec in each direction, 2GBytes/sec total.

I'm being an idiot and thinking of the base signalling rate which is 10Gbit/sec, before 8/10 encoding...