Tera Computer Company

A Scalable Computer

Index

Any computer manufacturer will claim their product scales. By this they mean the theoretical peak rates scale. At Tera, scalability means that the performance of the customer's application scales as processors are added. Performance increases in proportion to system size up to the limit of problem parallelism. Thus, if an engineering application runs at 350 megaflops on a single MTA processor, it will run at 350n megaflops on an n-processor system providing the problem is large enough. The promise of such performance is the reason why so many attempts have been made to port important codes like SPICE, DYNA3D and NASTRAN to existing parallel architectures. As the user community well knows, very few codes of interest scale on current architectures. Tera's breakthrough in performance is due to its multithreaded architecture, which removes the obstacles to scalability that plague competing systems.

Tera foresees the ability to build MTA systems that can provide sustained performance more than 100 times that of today's Cray T90/32 on codes like SPICE, DYNA3D and NASTRAN. The following chart shows how an initial Tera MTA-32C system installed at a customer's site can grow to meet future requirements. Tera chooses to illustrate performance in terms of application performance rather than peak rates. We believe that most users are unlikely to obtain greater than a 2.5-fold speed-up in performance on certain engineering applications with irregular grids no matter how many T90 CPUs are thrown at these problems. A similar lack of scaling for these types of problems is observed also on other VPPs and on microprocessor-based SMPs.


Anticipated Application Performance:


Year Proposed Tera
Configuration
Multiples of Cray T90/32
Performance
1997 MTA-32C 3.84 - 12.8
1998 MTA-64C 7.68 - 25.6
1999 MTA-128C 15.36 - 51.2
2000 MTA-256C 30.72 - 102.4

Our performance assumptions are based on the following:

  • The application is large and parallelized.
  • Sustained performance of each MTA processor will be between 150 and 500 megaflops (peak performance is 1 gigaflop).
  • Sustained Cray T90 performance will be 500 megaflops per processor (peak performance is 2 gigaflops).
  • The T90 parallel speed-up is 2.5 on typical engineering aplications.

Next:

[Tera's Solution] [A Scalable Computer] [A Programmable Computer] [A Parallel Computer] [A Breakthrough in HPC] [Hardware Characteristics] [Software Characteristics] [Our Conclusion]

[About Tera] [The MTA] [Press Releases] [Investing] [Work at Tera] [Partners] [Library] [Related Sites]

Contact Tera