SGI® NUMAflex® Architecture
• Up to 24 TB globally addressable memory
• All memory in system addressable by any processor
• Best message passing performance too
Most PC users have learned already that they have more to gain by adding memory in their systems than by adding compute power—this is even more true in supercomputing. SGI Altix offers the industry's most scalable memory architecture, offering up to 24TB of globally addressable memory in a system. This is in stark contrast to clustered architectures where addressable memory is typically limited to 32GB.
 |  |
Systems with global shared memory allow access to all data in the system's memory directly and efficiently, without having to move data through I/O or networking bottlenecks. The impact can be dramatic. For example, the latency or 'wait times' for messages that must go through system I/O are on the order of 1000 to 10000 greater than communication within the memory domain. To overcome this, clusters without global shared memory must instead pass copies of data, often in the form of messages, which can greatly complicate programming and slow down performance by increasing the time processors must wait for data.
Why does this matter? Users can do more and get results sooner. Entire databases can be driven directly out of memory. Hardware and software costs are lower. System administrators and users spend less time feeding and tuning the cluster. Developers have new flexibility to choose any programming model for application scalability. And it's all standards-based Linux. Read more about how SGI's customers and partners have already proven the benefits of Global Shared memory in a variety of industries.
Message Passing Performance:
Latency and Bandwidth performance of common interconnect technologies
| Technology | Vendor | MPI latency usec, short msg | Bandwidth per link (unidirectional, MB/s) |
| NUMAlink 4 (Altix) | SGI | 1 | 3200 |
| RapidArray (XD1) | Cray | 1.8 | 20001 |
| QsNet II | Quadrics | 2 | 9002 |
| Infiniband | Voltaire | 3.5 | 8303 |
| High Performance Switch | IBM | 5 | 10004 |
| Myrinet XP2 | Myricom | 5.7 | 4955 |
| SP Switch 2 | IBM | 18 | 5006 |
| Ethernet | Various | 30 | 100 |
Comparison of Linpack system efficiencies in the November 2004 Top 500 list
|
