My head is rolling around many thoughts related to networking and high-speed computer interfaces.
I know that 10Base-T Ethernet used Manchester encoding for sending “symbols” over the wire (in this case, I think each pair of ones and zeroes was encoded, though maybe it was each bit individually). The problem is that it required the physical wire speed to be twice your data rate, so a 10Mbps network would run at 20MHz. An advantage is that it can fairly easily self-synchronize, reducing the need for expensive high-accuracy clocks. This may have also been used for FDDI at 100Mbps, which would require a network capable of 200MHz operation (though I think that must have just been the copper wire variant and not the optical version).
I sort of mentioned in a previous entry that Fast Ethernet (100Mbps) improved upon this by using a “four out of five” encoding, where symbols composed of 5 bits were used to represent 4 bits of actual data. I think this is borrowed from some version of FibreChannel.
FibreChannel itself is descended from HIPPI, FDDI, and SCSI. HIPPI was originally a crazy interface used for linking supercomputers together. It used cables with 50 wires, and operated at very high speeds at the time. I think they had lots of problems with crosstalk, though. Crosstalk (where the signals from neighboring wires start to get blended together) is becoming an issue even on lower-end hardware, and is one of the big reasons we are now seeing a move to Serial-ATA in desktop machines.
In the big-ass hardware arena (supercomputers, datacenters, and Internet backbones), I've heard that SONET and InfiniBand are big. I believe SONET is primarily a network interface, and last I heard it was the speed champ, maxing out at 40 Gigabits per second. I don't know much about InfiniBand, but it seems to be more of a storage and system interconnect, and it appears to my untrained eyes to be very similar to FibreChannel. One interesting thing about it is that it appears to gang together multiple interfaces to achieve insane speeds. Connect together 12 channels each running at 2 Gigabits or so, and have fun.
Another curiosity I'm having at the moment regards Firewire. A new version just came out that doubles the speed. Originally, there was the IEE-1394 standard, circa 1995. I'm not sure exactly how fast that went, but it was either 100 or 200Mbps. 1394a, what is most common at the moment, runs at 400Mbps. 1394b has just started showing up, and is commonly being referred to as Firewire 800 because of it's 800Mbps speed. The strange thing about it is that Firewire 800 has a new connector, which I think is a very bad thing. Significant numbers of weird connectors has had a bad effect on SCSI, so making a new connector is problematic...
Another thing that caught my interest: Slashdot is saying that Apple may show off their next-generation “G5” desktop systems soon. Rumor has it that these systems will use IBM's 64-bit PowerPC 970 processor, and may use HyperTransport, which has been a fairly mythic bus architecture for some time. HyperTransport was originally designed for use with the (formerly DEC, then Compaq, and now, uh, nobody) Alpha processor, and I think AMD has looked at it heavily (they borrowed a lot of technology from the Alpha, which explains why Athlons have largely done so well).
It's becoming clear that personal computers will soon be moving to a new system bus. The old standard PCI interface will hang around a long time (nobody needs their sound card or 56k modem to be on an ultra-fast bus), but something new has to come in to handle high-bandwidth devices. We already have AGP for video cards, and the 64-bit PCI bus helps a lot too. Intel has the PCI-X interface, which runs 64-bit at 133MHz. I seem to recall hearing that HyperTransport will somehow be in competition with PCI-X, but I think they're two different things.
Clearly, this requires more research ;-)
Posted by mike at June 8, 2003 06:34 PM | Hardware | TrackBack