Silicon Graphics Inc has upgraded its Cray T3E-1200 supercomputer line with a new router chip that will also find its way into the company’s next generation of SGI Origin multiprocessors. The new chip, a custom ASIC, provides the new T3E-1200E massively parallel supercomputer with an aggregate global bandwidth of 42 gigabits per second, making larger configurations run more efficiently. Cray T3E machines can theoretically house up to 2,048 Alpha RISC processors, each with a clock speed of 600MHz. Pricing prohibits machines as large as that, but SGI says it does have four systems with over a thousand processors out at customer sites. The new router chip copes with faster and more numerous chips, and with larger and faster memory, by increasing measured interprocessor bandwidth from 330 to 420 megabytes per second, 25% up on its predecessor. It is also the key to the convergence of the T3E line with SGI’s Origin servers, which currently scale up to 128 processors. Although the two lines currently use different processors, the T3E’s distributed cache coherent architecture is very similar to the ccNUMA-based Origin, says SGI. They will come together in the next generation of machines. Beyond that, the scaleable vector line, currently represented by the Cray SV1, will be integrated into the converged T3E/Origin lines as a vector option. Meanwhile, SGI says it’s found that there are two sweet spots in supercomputer sales – systems of up to 256 processors, in the $8m to $10m price range, and very high-end systems, in the 1,000 processor class. SGI says it’s shipped two 1200E systems already, one to the CINECA research center in Bologna, Italy, and is currently installing a third at the UK Research Council in Manchester. It has four more orders for the new machine, a total of more than $70m in all.
