US tax payers shovel $10m into Cray R&D


ComputerWire: IT Industry Intelligence

Cray Inc, the Seattle, Washington supercomputing maker, announced this week that the U.S government has ponied up $10m to fund half of the development costs of future supercomputers based on the SV design,

Timothy Prickett Morgan writes


As part of the agreement, Cray has agreed to set aside an equal $10m for the research and development of the SV family of vector supercomputers.

The SV2 supercomputers designed by Cray (which was eaten in April 2000 by Tera Computer after being spun out of Silicon Graphics) are one of the key pillars of national security in an increasingly spooky and paranoid world.

In September 1999, Uncle Sam cut a check for an unspecified amount of development support for the SV2s, when Cray was still a unit of SGI, to make sure that the machine came to market. The government has been a big user of vector-based supercomputers - particularly at the National Security Agency, the Central Intelligence Agency, and the various branches of the military - and it looks as though the U.S. government wants to ensure that at least one domestic supplier of parallel vector supercomputers is always around to support its vector application needs.

Cray's statements on the deal would have us believe that Cray alone is dedicated to developing high-bandwidth, high-performance supercomputers and that is why Uncle Sam is paying $10m in development for future SV-style machines, but this is silly. IBM Corp, Hewlett Packard Co, SGI, and Sun Microsystems (to a certain extent) are committed to the HPC market and sure want to get a bite out of the ballooning military and intelligence budgets in the States. What Cray has and these companies do not are vector-style machines rather than massively parallel machines, and that means they run certain workloads better than can be done on parallel computers.

The U.S. government wants a variety of styles and architectures in indigenous supercomputers because that is what the workloads among its research institutions, its military branches, and its intelligence branches require. IBM, HP, Sun and to a certain extent SGI do not need development money from the U.S. government because they have very large general purpose server businesses on which they can shoulder the costs of parallel supercomputer development.

Cray doesn't have this, and this is why it needs Uncle Sam to pay forward on the machines it will want to buy years from now. None of this detracts from the importance of keeping Cray in business as a supercomputer vendor so there is an alternative to the parallel designs from other suppliers. This deal just illustrates that Cray has to be cultivated and tended in a more conscious manner than a free and brutal market might otherwise allow.

The SV line of vector machines have been around for almost three years. The SV1, SV1ex, and SV2 supers uses a mix of standard vector processors plus virtual vector processors, called Multi-Streaming Processors (MSPs). The SV1 line was launched by Cray when it was part of SGI in September 1999. Cray delivered the SV1ex line in the first quarter of 2001, using vector processors running at 450MHz, up from 300MHz with the SV1s.

The SV1 and SV1ex machines couple four CMOS processors together to create an MSP, which together are rated at a 7.2 gigaflops peak performance. In real world settings running vector programs, it comes in at between 3.6 gigaflops and 4.5 gigaflops, says Cray. The SV2s, which are expected later this year, are expected to have CMOS vector engines with about 3 gigaflops of power.

James Rottsolk, Cray's chairman and CEO, says that the $20m in funds will be used to enhance the processors and memory subsystems in the existing SV2 machines and to develop a successor to the SV2 that includes vector registers, distributed shared memory, higher memory bandwidth, stride-independent memory bandwidth (in short, this seems to mean improving how caches work with the processor), adding instruction-level parallelism to the SV processors, and making the overall machines burn less electricity and generate less heat.

Rottsolk that the SV2 can hit about 150 teraflops of peak performance with more research and development, up from about 50 teraflops of peak performance with the machines expected later this year. This is the first time Cray has put a number on the peak performance of a fully loaded SV2, something the company has been hinting at but not committing to.

Cray has not released any configuration details on the SV2 machines, but a machine with 4,096 MSPs, each delivering 12 gigaflops of power, yields 49 teraflops of aggregate peak processing capacity. In any event, the nearly 50 teraflops rating Cray is saying it can deliver with the SV2 puts the machine, in theory, ahead of the world's current most powerful supercomputer, NEC's Earth Simulator, which has 5,120 vector engines that collectively deliver around 40 teraflops peak computing power. Rottsolk says that with continued funding, the SV2 successor could handle several hundred teraflops and during its lifecycle should exceed one petaflop of peak power.

The announcement that Cray had received development funding from the U.S. government on the SV line of vector supers came hot on the heels of an announcement two weeks ago by Cray that it had been able to get in on the "Red Storm" 100 teraflops massively parallel supercomputer project being sponsored by the U.S. National Nuclear Security Administration, a division of the Department of Energy that helps fund the Sandia National Laboratories.

Sandia is said to be working out the finishing touches on a $90m contract with Cray to build such a machine, and already has a contract with Compaq Computer Corp from January 2001 along with Celera Genomics to build a 100 teraflops machine under the Red Storm supercomputer program. The Cray deal is an alternative supply contract, and this is consistent with the past practices of the U.S. government, which wants to foster innovation and competition among indigenous supercomputer suppliers.

Cray has also recently been awarded a one-year, $3.6m grant by the Defense Advanced Research Projects Agency's High Productivity Computing Systems program, which is asking the major supercomputer makers in the U.S. to come up with ideas about how to build the multi-petaflops supercomputers that the U.S. government will want to have in its arsenal in the 2009 to 2010 timeframe. IBM, Sun, and SGI were awarded similar contracts.

© ComputerWire

Biting the hand that feeds IT © 1998–2017