HyperTransport in High Performance Applications

The essence of today's high performance computing platforms are clusters of high performance processor or processing systems linked together using high bandwidth interconnects such as InfiniBand or 10Gb Ethernet. A key factor in system level performance is achieving low latency between the processors and clusters.

By using multiple processors with HyperTransport links, developers of high performance computing clusters can provide very high performance systems that are easily scaled upwards in performance. A processor cluster of four processors, each with its local caches and memory can utilize a HyperTransport switch to connect to each other and shared memory. Each processor has the full bandwidth available for accessing its own memory and can use the high-speed HyperTransport link to access the memory of the other processors.

As shown in the diagram, such systems are easily scaled to support many processor clusters. Each cluster can use HyperTransport links to connect to other processor clusters or to high-speed I/O channels or legacy PCI-type I/O devices.

Massive HyperTransport-based systems can be easily configured with multiple multiprocessor clusters linked together. Where additional bandwidth is required, 32-bit HyperTransport links can be used with 8- and 16-bit links used where the bandwidth needs are less.

For such systems, access to high-speed communications technologies such as 10 Gigabit Ethernet, InfiniBand, and SPI-4 can be accommodated using HyperTransport bridges to those technologies. HyperTransport technology's 22.4 Gigabit/second aggregate bandwidth enables it to easily integrate with these high-speed communications protocols.