Intel QuickPath Interconnect

QuickPath Interconnect (short: QPI ) is a technology developed by Intel point-to -point connection for communication between the processors with each other and for the communication between processors and chipset. QPI is in contrast to his predecessor Front Side Bus ( FSB ) is not a bus, but a routing mechanism that ensures that data packets reach the correct recipient CPU in a network of multiple processors. QPI was developed as a Common System Interface (CSI ), replacing the Core i7 and the associated introduction of the Nehalem architecture, the FSB, Intel -based systems from.

Technical

QPI has been designed to compete with AMD's HyperTransport technology and the first version is built since the second half of 2008.

A QPI port consists of two unidirectional QPI links, one for each direction of transmission. A QPI link consists of 20 line pairs for symmetrical transmission of 20 bits. QPI the first version is clocked at 2.4 and 3.2 GHz and operated in the DDR process with two transfers per clock cycle or a transfer per clock edge, giving a rate of 4.8 or 6.4 giga per second ( GT / s ) results. A link with 20 line pairs provides 20 bits per transfer, of which 16 bits are used to transmit data and 4 bits for header information. The bandwidth of the QPI links can be calculated as follows:

At 2.4 GHz and 4.8 GT / s:

At 3.2 GHz, or 6.4 GT / s:

As QPI is a bidirectional point-to -point connection, two ( one-way ) links are used in a QPI connection. Therefore, the above ranges of links are multiplied by two. A QPI link has a maximum bandwidth of 19.2 GB / s and 25.6 GB / s Compared with the fastest frontside bus, FSB 1600, which provides a transfer rate of 12.5 GB / s QPI the first version twice as fast (4 × 400 MHz × 8 bytes (64 bits): 1024 = 12.5 Gb / s).

However, the consideration of the bandwidth alone is not meaningful for the speed of a system. Add to that the latency to be as important criterion. When FSB is probably less because it is a bus system and no routing takes place here. The cross-bar in the new router Nehalem processors is an additional component to the transmission path, which increases the signal propagation times, and therefore the latency, which does not affect the data rate, the speed of the overall system but reduces. What in practice is faster now, FSB1600 or QPI the first version, depends on the application. If the number of consecutive transmit small amounts of data, FSB is due to the lower latency advantage for large amounts of data at a time QPI. Since QPI can transfer bidirectional, the disadvantage of the higher latency by the FSB is balanced if the application requires it, which is certainly the case of multi -core processors is the case if, for example, a core fetches data from memory and the other simultaneously transmits data to the graphics card.

The transmission on a QPI link can be found in 80 - bit chunks instead ( Intel flit called ), for which ( called phits ) at 20 line pairs per one bit 4- transfers and transfer cycles are needed. Of the 80 bits of a flit is 64 bits ( 4 x 16-bit) user data, and the remaining 16 bits (4 x 4 bit) header values ​​. Header values ​​are used by the link-layer for CRC error correction, Message Classes and Virtual Channels.

Until March 31, 2010, QPI clock speed of 3.2 GHz was only reserved for the larger Core i7 models. On this day, however, Intel has announced that the Core i7 920, and the other i7 CPUs with a QPI of 4.8 GT / s, without loss of 6.4 GT / s of the guarantee can be operated. This applies to all the models, including the older C0 stepping.