Slashdot Mirror
PCI Express - Coming Soon to a PC Near You
Max Romantschuk writes "I've been following the emerging of PCI Express for some time now. PCI Express, previously known as "Third Generation I/O" or "3GIO", is the technology set to replace PCI. PCI has been with us for around ten years now, and is rapidly running out of bandwidth. Last week Anandtech ran an interresting story on PCI Express. The techology has previously been covered by Hexus and ExtremeTech aswell. I feel this technology looks all set to replace PCI, and we really do need some new bus technology to keep up with the bandwidth demands of today's applications. Or is this just yet another way to force us into a new upgrade cycle?"
Due to its high bandwidth, it's expected to replace AGP as well.
"Oooh, does that mean we get to kick some puffy white mad zionist butt?"
As I understand it, using serial there is no having to worry about whether all the bits arrive at the same time (as there obviously is with parallel), and so the speed of transmission can be dramatically increased past the point at which it becomes faster than the "equivalent" parallel technology... bits arrive in the order they were sent - guaranteed.
University - a box of academia nuts.
The limited number of IRQs hasn't been a problem since PCI 2.1 and APIC. It's a problem with Windows 9x and a few other operating systems, but those won't be able to use PCI Express anyhow.
The problem today is more with interrupt line sharing (#A, #B, #C, #D -- some motherboards add more, but four is the old spec), and cards sharing the actual interrupt and not the interrupt queue (IRQ), depending on how you place them.
But yes, to answer your question, there's less problems, due to the parallel serial nature (now is that an oxymoron?) of the controller interface, working somewhat like SCSI does.
At least until 4x, 8x and 16x PCI Express arrives in force, and cards starts competing and assuming that all the streams are available for THEIR card, much like some cards today think it's ok to bump up the PCI latency, cause the user SURELY must have some unpopulated slots we can steal time from...
Yeah, when that happens, it may be hell to troubleshoot, but we'll just wait and see...
Now that's a no-brainer.
My computer is by far not a high-end box, but PCI is a (small) bottle-neck, even for me.
Let's see: 2 IDE channels, 2 disks, that's 50 MB/s each, 1 GBit network, that's peak 100MB/s. A U2W SCSI host adapter with 1 single, very fast disk is good for 70MB/s. Then there is USB2 (everything is USB2 now) and Firewire (each 50MB/s). Adds up to (peak) 370MB/s.
You and me and most people know, that a usual user and most unusual users like the ./ crows will never use all devices at once. But just copying data from disk to network saturates the bus.
A simple fix is 66MHz 64Bit PCI, but those are very rare in consumer machines. So while PCI-Express might be currently overkill, I doubt simple 33MHz 32Bit PCI will be sufficient even for consumer grade computers. Just imagine 10 years ago when PCI started: most were using ISA and that was enough for most usual users. 10MBit/s Ethernet cards used less than 1MB/s. Who needs a faster bus? Only servers needed PCI (or EISA).
Watching the long migration from ISA to PCI until ISA was (mostly) replaced, I don't expect PCI-Express to replace PCI within 5 years. And in 5 years I would bet, that PCI looks like ISA does now: slow and outdated.
using serial there is no having to worry about whether all the bits arrive at the same time (as there obviously is with parallel), and so the speed of transmission can be dramatically increased past the point at which it becomes faster than the "equivalent" parallel technology... bits arrive in the order they were sent - guaranteed.
:-l
I'm afraid this might add to the confusion about serial interfaces being 'faster' than parallel. While it is true that you don't have to worry about data/clock skew when using serial interfaces, enabling you to clock them faster, a parallel interface running at the same clock speed as a serial interface will always be faster in terms of data throughput. The reason for this is simple: serial == 1 bit per clock, parallel = > 1 bit per clock.
So, saying that serial is faster than the "equivalent" parallel interface is confusing, and incorrect, because one could be referring to equivalent clock rates being used for each interface, in which case parallel will provide at least twice the data throughput. On the other hand, "equivalent" could be referring to identical throughput rates, in which case the serial and parallel interfaces would provide, by definition, identical data rates.
The real advantage that PCI Express has over PCI/PCI-X is that it is a point-to-point, rather than a multi-drop, bus. This setup requires less time between pin transitions, meaning that it can be clock faster. Also, like Ethernet, a serial protocol can imbed the clock into the data stream so clock/data skew is no problem whatsoever.
Serial is not better than parallel anymore than digital is better than analog, there are just physical reasons why implementing point-to-point serial at significantly higher clock rates is easier than multi-drop parallel.
Anyone still awake?
Didn't think so
If it's not one thing, it's Steve's Mother
We realized PCI wasn't going to be fast enough years ago-- that's why pretty much every motherboard you can buy today has an AGP socket.
And even that wasn't fast enough, now we have AGP 8x.
But seriously, is PCI really not fast enough for the general consumer, once he's got an AGP socket? PCI that runs on a 66MHz bus that's 64 bits wide has existed and even been available in high-end PC class hardware for years, but few of even Slashdotters have anything other than 32 bit 33MHz PCI in our home machines. The only time I ever deal with the 64 bit PCI cards is for Sun Microsystems hardware at the office.
I don't think this is "forcing another upgrade cycle" at all-- upgrades already exist, and most of us don't have 'em.
The one consumer-level application where a better bus is vital is HDTV. Current buses are just barely able to handle uncompressed SDTV (20-30MB/sec). (in theory PCI gives you 133MB/sec and AGP more, but as they say, in theory there is no difference between theory and practice :).
PCI-X will finally bring HDTV (~200MB/sec) within reach. What this means is that you'll be able to have a software-only HDTV decoder - which will make it trivial to receive HDTV broadcasts on a PC, and make HD-DVD players possible.
At the pro level, this is just about the last thing that a $50K SGI system has over a cheap Linux PC - playback and maniupulation of uncompressed HDTV video. It's about time PCs finally caught up to "workstations" in the bus department...
Moderators on crack. This is just plain wrong. Firewire is 400 or 800 Mbits/s, while SCSI is up to 320 Mbyte/s, IDE is up to 133 MByte/s, and Fibre Channel is up to 250 Mbyte/s. These numbers are directly comparable, because different buses have different amounts of overhead, but for sure firewire is a slow also-ran when talking only about performance. (When talking about cost, flexibility, etc., firewire looks better, of course.) As far as PCI goes, the top end is over 1 Gbyte/s, which is a bottleneck for some applications, but not firewire. Also, in high-end servers you'll have a number of pci buses to improve performance.