Technology

The Cray XMT system was architected to leverage Cray's MPP system design to create a scalable, reliable and economical multithreaded supercomputing platform. The design is based on a Cray MPP compute blade but utilizes AMD Torrenza technology to populate the AMD Opteron sockets with custom Cray Threadstorm™ chips developed for multithreaded processing. A single Cray Threadstorm processor can sustain 128 simultaneous threads and is connected with up to 8 GB of memory that is globally accessible by any other Cray Threadstorm processor in the system.

Each Cray Threadstorm processor is directly connected to a dedicated Cray SeaStar2™ interconnect chip, resulting in a high bandwidth, low latency network characteristic of all Cray systems. This allows the Cray XMT platform to scale to a system with over 8000 processors providing over one million simultaneous threads and 64 terabytes of shared memory.

As another technology using the Cray XT™ infrastructure, the Cray XMT platform includes separate AMD Opteron-based service blades that can be configured for I/O, login, network or system functions and can also provide scalar processing for applications that are best served by a combination of scalar and multithreading technologies.