Data sets in the enterprise are growing larger everyday, and the move towards 64-bit computing will exacerbate the trend. Even with today’s 32-bit computer systems, the bottleneck for data is no longer the CPU, but the interconnection points — buses — between the system components.
To reach the high rates necessary to handle large data volumes, manufacturers are increasingly moving towards channelised serial protocols, which scale upwards better than their parallel counterparts. Using serial protocols also means easier physical implementation, with thinner cables and smaller connectors that are cheaper to make and do not impede airflow in computer cases.
Removable peripherals: USB 2.0 and IEEE-1394 (FireWire/i.Link)
Even though most computers still come with a parallel port (transfer data at 2MB/s to 3MB/s) and at least one serial port (115kbit/s), they are about to be replaced by much faster serial buses like universal serial bus (USB) and IEEE-1394 FireWire.
USB is the more common of the two and is used for a great variety of devices including modems, uninterruptible power supplies, mice, keyboards, scanners and cameras. There are even USB network adapters, removable hard disks and memory card storage devices, and you can find USB hubs in monitors, for easy access.
While the first generation of USB devices was limited to a maximum of 12Mbit/s bandwidth shared with all devices, version 2.0 enables some 480Mbit/s signalling speed, making it fast enough for digital media devices like cameras.
The low price of implementation and the ease of use (devices are often hot-pluggable and auto-configuring) has made USB popular with vendors and customers alike; furthermore, new computers come with four or more USB ports, allowing for many devices to be connected simultaneously — by using USB hubs, the theoretical number of devices per bus is 127.
Soon, devices equipped with USB On-The-Go ports will be able to communicate not just with PCs but also with each other. That is, your mobile phone can be plugged into a printer, or an MP3 player or a mass storage device, for instance.
Even though it was initially faster, operating at 400Mbit/s, the IEEE-1394 serial bus hasn’t had the impact of USB and isn’t usually bundled with PCs but with recent Apple Macintosh computers. Typically, IEEE-1394 devices are at the higher end of the price scale than USB ones, like camcorders and professional digital still cameras.
IEEE-1394 costs more to implement than USB and the current standard has a limit of 63 simultaneously connected devices and a maximum cable length of 4.5m for the full 400Mbit/s signalling (14m for 200Mbit/s). However, just like USB, IEEE-1394 is hot-swappable and has had peer-to-peer capability since the beginning, like USB OTG now offers.
But whereas USB 2.0 looks like the end of the road for the technology, with no further revisions on the horizon,. IEEE-1394 is evolving, with the “b” specification finalised last year to include speeds of 800, 1600 and 3200Mbit/s to allow for future scalability.
Mass storage: Serial ATA and serial attached SCSI
For mass-market PCs and low-end business servers, serial ATA (SATA) will replace parallel ATA (PATA) as the bus for mass storage devices like hard drives.
SATA offers several advantages over PATA: it’s a serial bus using low voltages (250mV compared to 5V) and thin cables up to 1m in length that are easy to route, unlike the shorter, flat and wide PATA cables.
Master and slave jumpers are also a thing of the past — SATA is point-to-point, with one drive per channel in normal usage.
SATA is faster, with the 1.0 specification offering 1.5Gbit/s compared to 133MB/s (1Gbit/s) for PATA, with a roadmap specifying 3Gbit/s and 6Gbit/s over the next 10 years. (Note that SATA uses “8B10B” encoding, which effectively gives 10-bit bytes, not 8-bit ones.)
Although no ATA hard drive is able to churn out data at even 150MB/s, the added speed is welcome for RAID-0 and 1+0 configurations using multiple fast ATA drives for added performance. Furthermore, SATA storage can scale to terabyte level by using a Port Multiplier that allows for up to 15 devices per channel.
The parallel SCSI bus is still going strong, with devices rated at a maximum of 320MB/s currently. However, even SCSI is moving to a serial technology, termed serial attached SCSI (SAS), with the first devices due later this year and early 2004.
Just like SATA, SAS promises smaller connectors and thinner cables (up to 6m long), but is faster at 3Gbit/s and uses full-duplex wiring, allowing more than 128 devices on each channel. Drives will be dual-ported instead of single-ported like on SATA, and the next revision increment promises speeds of 6Gbit/s.
SATA devices can be connected to SAS adapters, providing the ability to set up cheap JBOD (just a bunch of disks) farms of hot-swappable ATA drives, with mission-critical data on more expensive SAS SCSI drives, all in the same enclosure.
SAS promises to deliver the performance of fibre channel arbitrated loop (FCAL) storage that signals at 2Gbit/s currently, but at a lower cost.
Internal expansion: PCI Express
The venerable 32-bit, 33MHz peripheral component interconnect bus’s days are numbered. With only a theoretical 132MB/s of bandwidth available for all the devices on the PCI bus, the technology isn’t sufficient to run gigabit ethernet expansion cards at full rate. Furthermore, it is 32 bits short of a full 64, to complement the next generation of CPUs.
The industry is split as to what will replace PCI, however. The backwards compatible PCI-X standard, which runs at 66/133MHz and comes in 32/64-bit versions providing a maximum throughput of 1GB/s has already been deployed in servers and high-end workstations.
Next up, PCI-X 2.0 adds two new speeds, 266 and 533MHz. This lifts throughput to some 2GB/s and 4GB/s respectively in 64-bit mode, a whopping 32 times as fast as the original PCI bus, without changes to the physical connectors on motherboards, or backwards compatibility with older devices.
PCI-X 2.0 is supported by heavy-hitters like server chip set maker ServerWorks as it offers a natural, evolutionary path towards increased performance.
Even so, the PCI special interest group (SIG) has been working on a radically different successor to the original PCI standard. Formerly known as 3GIO (third-generation I/O), it was renamed PCI Express last year, and is a serial technology that promises extremely high speeds, but does not retain backwards compatibility with existing PCI cards.
The initial release of PCI Express runs at 2.5GHz on an 0.8V bus, providing some 250Gbit/s per bidirectional serial channel (or lane). Combining the lanes to a maximum of 32 gives a grand total of 16GB/s of symmetric bandwidth.
One advantage of PCI Express is that it can be scaled using both the number of lanes as well as the clock frequency, meaning you could use single or two-lane interconnects for devices that do not require massive amounts of bandwidth (such as legacy PCI cards, for instance). The single and dual-lane connectors are much smaller than the sixteen and thirty-two lane ones, saving valuable motherboard real estate.
Interestingly enough, PCI Express connectors will also be used for PCMCIA cards — code-named “New Cards”, they will contain not just PCI Express, but also USB 2.0 and the systems management bus. Without PCI Express, mobile users would have to rely on built-in gigabit ethernet interfaces, as the current PCMCIA specification is only 32-bit and runs at 33MHz or the same as PCI.
Later this year, PCI Express will replace the Advanced Graphics Port (AGP) bus for video cards, with a 16-lane, 8GB/s version. This compares to 2.1GB/s for the current AGP 3.0 8X specification.
PCI Express video cards have already been announced by workstation graphics specialist 3DLabs. Other large vendors such as ATI and NVIDIA as well as Intel have also expressed support for PCI Express as the future interconnect standard, so it is the PCI replacement technology that most users will see first as it will appear in desktops, not servers like PCI-X.