Torrenza, launched in June 2006, is a specification for connecting a co-processor, such as a graphics chip or application-specific processor, directly to the HyperTransport interface in AMD's Opteron microprocessor. The connection is used to boost overall performance in compute-intensive applications.
Torrenza is used primarily in high-performance computing found in scientific research. AMD, however, believes it has "gotten over the hump" in pushing the technology into more commercial uses, Doug O'Flaherty, division manager for acceleration strategies at AMD, told InformationWeek in an interview from the conference.
While not claiming a dramatic adoption rate, AMD is seeing broader interest in the technology, O'Flaherty said. While the company doesn't have any hard numbers to show progress, AMD is pointing to examples, such as Activ Financial Stystems' use of Torrenza. The company is moving its in-memory database to combined AMD processor with a field programmable gate array -- a semiconductor device that contains programmable logic components, and programmable interconnects.
Hardware built with Torrenza can talk to an AMD processor either through a chip socket on the motherboard or through a PCI Express slot, which is a high-speed peripheral interconnect. In another example of a Torrenza-based commercial application, Hewlett-Packard plans to use the specification in connecting the HTX cards in its upcoming ProLiant Servers with AMD's new quad-core Opteron processors. The cards enable customers to deploy customized acceleration technology in their data centers.
Besides financial institutions, AMD says Torrenza-related technology will eventually find its way in products sold to the oil and gas industry and life sciences, and in commercial applications that perform large amounts of data manipulation, such as in indexing or search.
To help in building software that takes advantage of Torrenza-based devices, RapidMind plans to release development tools for FPGAs, O'Flaherty said. "We're trying really hard to span software, hardware deployments." RapidMind currently provides tools for programming to AMD and Intel multi-core CPUs, IBM's Cell processor, and Nvidia graphics processors.
AMD's Torrenza is battling with Geneseo, set of extensions to PCI Express trumpeted by Intel and IBM to define interconnects for their multi-core processors. The idea is to build networks of partners offering co-processors that can plug into the vendors' CPUs for faster computing.