P55 DMI>CPU bandwidth?

[2senile]

Can anybody tell if there is a problem with bandwidth from the PCH to CPU on the new Lynnfield motherboards?

I originally spotted this article about an issue with SATA 6GB/s.

http://www.behardware.com/news/10424/red-flag-on-current-sata-6-gbit-s.html

The first problem is that the only SATA 6 Gbit controller available, the Marvell 88SE9123, only connects with one PCI-Express link. At best then, you get 500 MB/s, which is less than the theoretical speeds of SATA 6 Gbit/s, without counting the fact that this bandwidth will have to be shared between the two ports supported by the chip. At worst then, you’ll have 250 MB/s per port!

The second problem is that the additional PCI-Express lanes on the Intel chipsets don’t work at 5 GT/s but at 2.5 GT/s, even on the latest P55 Express. A simple design linking the Marvell 88SE9123 to the P55 Express will therefore limit it to 250 MB/s, which is less than the bandwidth you get with SATA 3 Gbit/s!

From looking at the block diagram for the P55 chipset it appears the DMI link is 2GB/s.

Checked an Intel PDF.

Direct Media Interface
— 10 Gb/s each direction, full duplex

 

[pcchen]

Intel decided to limit the PCI express lanes on the P55 chipset to 1.0 speed (they support 2.0 but only 1.0 speed), probably because the DMI is not fast enough (1GB/s for each direction is only good for 4 PCI express lanes @ 250MB/s for each direction). So the Marvell chip which only supports one PCI express lane can only connect to the chipset with 250MB/s speed for each direction.

If you don't need 6Gbps SATA (which, if you use Marvell chip for that, you can't get 6Gbps anyway) then P55 has native support for 3Gbps SATA ports, which should be good enough for most applications. However, the total bandwidth between P55 and the CPU is still limited to 1GB/s per direction though.

 

[2senile]

Thanks for the reply, really appreciate it. I'm afraid I'm not as technically adept as I might wish & I was thinking I must have been missing something.

Is there any danger of overwhelming that 1GB/s with a couple of SATA II disks in Raid 1, an SSD, TV Tuner/capture & X-Fi soundcard? For example?

 

[mczak]

Is there any danger of overwhelming that 1GB/s with a couple of SATA II disks in Raid 1, an SSD, TV Tuner/capture & X-Fi soundcard? For example?

Soundcard shouldn't require a lot of bandwidth, TV capture neither (though it depends I guess analog hdtv would be a lot of bandwidth, but even full dvb stream with multiple channels doesn't exceed 10MB/s). Gb ethernet also can use more than 100 MB/s. I guess you could saturate the DMI link with something like 4 SSDs, though I'm not sure if the integrated sata controller can actually use that much bandwidth itself or is limited to less.

 

[mczak]

Intel decided to limit the PCI express lanes on the P55 chipset to 1.0 speed (they support 2.0 but only 1.0 speed), probably because the DMI is not fast enough (1GB/s for each direction is only good for 4 PCI express lanes @ 250MB/s for each direction).

I don't really see the problem. AFAIK DMI itself is pretty much pcie x4 connection, so if intel boosted pcie speed to 2.0 levels (and the pcie ports for pcie integrated into the cpu are already faster anyway) it would only be natural that DMI would get twice as fast too.
Naturally, you couldn't use older ICHs with boosted DMI, which at least in theory could work otherwise (I am actually wondering about this, since reportedly PCH prices are high why some motherboard vendors didn't opt for ICH10 or older instead even if not officially sanctioned - guess you can't actually buy only ICH from intel without MCH, but at least from a functionality point of view you wouldn't actually lose anything you'd miss by using ICH10R instead of P55).

If you don't need 6Gbps SATA (which, if you use Marvell chip for that, you can't get 6Gbps anyway)

well, the chip may only have 1 lane, but it's possible to connect it to the chipset via a pcie bridge (using pcie x2/x4 1.0 at the other end) to get faster performance (I think asus wanted to do that with an extension card?). Not an economical solution I agree, but otherwise the sata III chip is indeed completely useless.

 

[Grall]

Why worry over SATA3? Raw data transfer speed is the least important metric in what determines a fast data storage subsystem. Access time, IOPS are vastly, vastly more important.

You could connect today's fastest SSDs over sata1 links, and they would perform pretty much as fast in almost all situations...

It's no big deal. Wait until sata3 becomes properly integrated into chipsets, there won't be very many devices supporting the standard until then anyway.

 

[2senile]

Thanks to everybody for their thoughts on this.

I think I'll just go back to the original full system upgrade plan I had before the i750 was released.

 

[ShaidarHaran]

Why worry over SATA3? Raw data transfer speed is the least important metric in what determines a fast data storage subsystem. Access time, IOPS are vastly, vastly more important.

You could connect today's fastest SSDs over sata1 links, and they would perform pretty much as fast in almost all situations...

It's no big deal. Wait until sata3 becomes properly integrated into chipsets, there won't be very many devices supporting the standard until then anyway.

Depends on the usage pattern. If your access time is incredibly low and so is your throughput, it will take longer to transfer large amounts of data than if your access time is mediocre but your throughput is excellent.

 

[Grall]

SATA "1" throughput isn't "incredibly low" though. It's actually roughly as fast or even faster than even the fastest mechanical 7200RPM consumer disks out there. For example, the new Seagate (ironically SATA 3 compatible) 500GB/platter 2TB unit is specced as just under 140MB/s maximum transfer speed. This is for the outermost section of the platters, obviously. Inner cylinders would be slower, by a lot...

Raw throughput isn't really a bottleneck even for original SATA unless you sit there and copy a lot of rather big files.