[ Team LiB ] Previous Section Next Section

HyperTransport

The fastest way to send signals inside a computer is via HyperTransport, but HyperTransport is not an expansion bus. It isn't a bus at all. Strictly defined, it is a point-to-point communications system. The HyperTransport specification makes no provision for removable expansion boards. However, HyperTransport is destined to play a major role in the expansion of high-performance computers because it ties the circuitry of the computer together. For example, it can link the microprocessor to the bus control circuitry, serving as a bridge between the north and south bridges in the motherboard chipset. It may also link the video system to the north bridge. Anywhere signals need to move their quickest, HyperTransport can move them.

Quick? HyperTransport tops out with a claimed peak data rate of 12.8GBps. That's well beyond the bandwidth of today's microprocessor memory buses, the fastest connections currently in use. It easily eclipses the InfiniBand Architecture connection system, which peaks out at 2GBps.

But that's it. Unlike InfiniBand, which does everything from host expansion boards to connect to the World Wide Web, HyperTransport is nothing but fast. It is limited to the confines of a single circuit board, albeit a board that may span several square feet of real estate. As far as your operating system or applications are concerned, HyperTransport is invisible. It works without any change in a computer's software and requires no special drivers or other attention at all.

On the other hand, fast is enough. HyperTransport breaks the interconnection bottlenecks in common system designs. A system with HyperTransport may be several times faster than one without it on chores that reach beyond the microprocessor and its memory.

History

HyperTransport was the brainchild of Advanced Micro Devices (AMD). The Texas chip-maker created the interconnection design to complement its highest performance microprocessors, initially codenamed Hammer.

Originally named Lightning Data Transport, the design gained the HyperTransport name when AMD shared its design with other companies (primarily the designers of peripherals) who were facing similar bandwidth problems. On July 23, 2001, AMD and a handful of other interested companies created the HyperTransport Technology Consortium to further standardize, develop, and promote the interface. The charter members of the HyperTransport Consortium included Advanced Micro Devices, API NetWorks, Apple Computers, PMC-Sierra, Cisco Systems, NVidia, and Sun Microsystems.

The Consortium publishes an official HyperTransport I/O Link Specification, which is currently available for download without charge from the group's Web site at www.hypertransport.org. The current version of the specification is 1.03 and was released on October 10, 2001.

Performance

In terms of data transport speed, HyperTransport is the fastest interface currently in use. Its top speed is nearly 100 times quicker than the standard PCI expansion bus and three times faster than the InfiniBand specification currently allows. Grand as those numbers look, they are also misleading. Both InfiniBand and HyperTransport are duplex interfaces, which means they can transfer data in two directions simultaneously. Optimistic engineers calculate throughput by adding the two duplex channels together. Information actually moves from one device to another through the InfiniBand and HyperTransport interfaces at half the maximum rate shown, at most. Overhead required by the packet-based interfaces eats up more of the actual throughput performance.

Not all HyperTransport systems operate at these hyperspeeds. Most run more slowly. In any electronic design, higher speeds make parts layout more critical and add increased worries about interference. Consequently, HyperTransport operates at a variety of speeds. The current specification allows for six discrete speeds, only the fastest of which reaches the maximum value.

The speed of a HyperTransport connection is measured in the number of bits transferred through a single data path in a second. The actual clock speed of the connection on the HyperTransport bus is 800MHz. The link achieves its higher bit-rates by transferring two bits per clock cycle, much as double data-rate memory does. One transfer is keyed to each the rise and fall of the clock signal. The six speeds are 400, 600, 800, 1000, 1200, and 1600Mbps.

In addition, HyperTransport uses several connections in parallel for each of its links. It achieves its maximum data rate only through its widest connection with 32 parallel circuits. The design also supports bus widths of 16, 8, 4, and 2 bits.

Table 9.5 lists the link speeds as well as the maximum bus throughputs in megabytes per second at each of the allowed bus widths.

Table 9.5. HyperTransport Peak Throughput As a Function of Bus Width and Speed
  HyperTransport Link Width (In Bits)
Link Speed in Mbps 2 4 8 16 32
400 100 200 400 800 1600
600 150 300 600 1200 2400
800 200 400 800 1600 3200
1000 250 500 1000 2000 4000
1200 300 600 1200 2400 4800
1600 400 800 1600 3200 6400

Certainly, a wider bus gives a performance edge, but the wider connections also rapidly increase in complexity because of the structure of the interface.

Structure

HyperTransport is a point-to-point communications system. Functionally, it is equivalent to a wire with two ends. The signals on the HyperTransport channel cannot be shared with multiple devices. However, a device can have several HyperTransport channels leading to several different peripheral devices.

The basic HyperTransport design is a duplex communications system based on differential signals. Duplex means that HyperTransport uses separate channels for sending and receiving data. Differential means that HyperTransport uses two wires for each signal, sending the same digital code down both wires at the same time but with the polarity of the two signals opposite. The system registers the difference between the two signals. Any noise picked up along the way should, in theory, be the same on the two lines because they are so close together, so it does not register on the differential receiver.

HyperTransport uses multiple sets of differential wire pairs for each communications channel. In its most basic form, HyperTransport uses two pairs for sending and two for receiving, effectively creating two parallel paths in the channel. That is, the most basic form of HyperTransport moves data two bits at a time. The HyperTransport specification allows for channel widths of 2, 4, 8, 16, or 32 parallel paths to move data the corresponding number of bits at a time. Of course, the more data paths in a channel, the faster the channel can move data. The two ends of the connection negotiate the width of the channel when they power up, and the negotiated width remains set until the next power up or reset. The channel width cannot change dynamically.

The HyperTransport specification allows for asymmetrical channels. That is, the specification allows for systems that might have a 32-bit data path in one direction but only 2 bits in return.

Although HyperTransport moves data in packets, it also uses several control lines to help reduce software overhead and maintain the integrity of the network. A clock signal (CLK) indicates when data on the channel is valid and may be read. Channels 2-, 4-, or 8-bits wide have single clock signals; 16-bit channels have two synchronized clock signals; 32-bit channels have four clock signals. A control signal (CTL) indicates whether the bits on the data lines are actually data or control information. Activating this signal indicates the data on the channel is control information. The line can be activated even during a data packet, allowing control information to take immediate control. A separate signal (POWEROK) ensures the integrity of the connection, indicating that the power to the system and the clock are operating properly. A reset signal (RESET#) does exactly what you'd expect—it resets the system so that it reloads its configuration.

For compatibility with power-managed computers based on Intel architecture, the HyperTransport interface also includes two signals to allow a connection to be set up again after the system goes into standby or hibernate mode. The LDTSTOP# signal, when present, enables the data link. When it shifts off, the link is disabled. The host system or a device can request the link be reenabled (which also reconfigures the link) with the LDTREQ# signal. Table 9.6 summarizes the signals in a HyperTransport link.

Table 9.6. HyperTransport Bus Signals
Mnemonic Function Width
CAD Command, address, and data lines 2, 4, 8, 16, or 32
CTL Control 1
CLK Clock 1, 2, or 4
PWROK Power and clock okay 1
RESET# Reset chain 1
LDTSTOP# Enable/disable link 1
LDTREQ# Request link enable 1

The design of the HyperTransport mechanical bus is meant to minimize the number of connections used because more connections need more power, generate more heat, and reduce the overall reliability of the system. Even so, a HyperTransport channel requires a large number of connections—each CAD bit requires four conductors, a two-wire differential pair for each of the two duplex signals. A 32-bit HyperTransport channel therefore requires 128 conductors for its CAD signals. At a minimum, the six control signals, which are single ended, require six more conductors, increasing to nine in a 32-bit system because of the four differential clock signals.

Each differential pair uses a variation on the low-voltage differential signaling (LVDS) standard. The differential voltage shift between the pairs of a signal is only 600 millivolts.

Standards and Coordination

The HyperTransport specification is maintained by the HyperTransport Technology Consortium, which can be reached at the following addresses:

HyperTransport Technology Consortium

1030 E. El Camino Real #447

Sunnyvale, CA 94087

Phone: 800-538-8450 (Ext. 47739)

E-mail: info@hypertransport.org

Web site: www.hypertransport.org

    [ Team LiB ] Previous Section Next Section