Looking at All of the Options

Jerry Brost

Gerald M. Brost, Vice President of Engineering for Cray Research, Inc. (CRI), has been with the company since 1973 and has made significant contributions to the development and evolution of CRI supercomputers, from the CRAY-1 through the Y-MP C90. His responsibilities have included overall leadership for projects involving the CRAY X-MP, the CRAY-2, the CRAY Y-MP, CRAY Y-MP follow-on systems, and Cray's integrated-circuit facilities and peripheral products. Today, his responsibilities include overall leadership for the CRAY Y-MP EL, the CRAY Y-MP C90, the MPP Project, and future product development.

Before joining Cray, Mr. Brost worked for Fairchild Industries on a military system project in North Dakota. He graduated from North Dakota State University (NDSU) with a bachelor of science degree in electrical and electronics engineering and has done graduate work in the same field at NDSU.

To remain as the leaders in supercomputing, one of the things that we at Cray Research, Inc., need to do is continue looking at what technology is available. That technology is not just circuits but also architecture. We need to keep looking at all the technological pieces that have to be examined in order to put together a system.

Cray Research looked at technologies like gallium arsenide about eight years ago and chose gallium arsenide because of its great potential.

Today it still has a lot of potential, and I think someday it is going to become the technology of the supercomputer.

We also looked at optical computing and fiber optics, which is an area in which we will see continued growth. However, we are not committed to optical-circuit technology to build the next generation of Cray supercomputers.

Several years ago, we looked at software technology and chose UNIX because we saw that was a technology that could make our systems more powerful and more usable by our customers.

Superconductors look like a technology that has a lot of potential. However, we are unable to build anything with superconductors today.

It may come as a surprise to some that massively parallel architectures have been out for at least 20 years. Some people might say that these architectures have been out longer than that, but they have been out at least 20 years.

Even in light of the available technologies, are we at a point where, to satisfy the customers, we should incorporate the technologies into our systems? Up to now, I think the answer has been no.

We have gone out and talked to our customers and surveyed the customers on a number of things. First of all, when we talked about architectures, we were proposing what all of you know as the C90 Program. What should that architecture look like? We have our own proposal, and we gave that to some of the customers.

We talked about our view of massive parallelism. We asked the customers where they saw massive parallelism fitting into their systems. It is something that really works? Although you hear all the hype about it, is it running at a million floating-point operations per second—a megaflop? Or is it just a flop? One of the things that we learned in our survey on massive parallelism is that there are a number of codes that do run at two- to five-billion floating-point operations per second (two to five GFLOPS).

If I listen to my colleagues today, I hear that there are large numbers of codes all running at GFLOPS ranges on massively parallel machines. Indeed, there has been significant progress made with massively parallel machines. There has been enough progress to convince us that massively parallel is an element that needs to be part of our systems in the future. Today at Cray we do have a massively parallel program, and it will be growing from now on.

Massively parallel systems do have some limitations. First of all, they are difficult architectures to program. For many of the codes that are running at the GFLOPS or five-GFLOPS performance level, it probably

took someone a year's time to get the application developed. But that is because the tools are all young, the architecture is young, and there are a lot of unknowns.

Today there are probably at least 20 different efforts under way to develop massively parallel systems. If we look at progress in massive parallelism, it is much like vector processing was. If we go back in time, basically all machines were scalar processing machines.

We added vector processing to the CRAY-1 back in 1976. At first it was difficult to program because there were not any compilers to help and because people didn't know how to write special algorithms. It took some time before people started seeing the advantage of using vector processing. Next, we went on to parallel processors. Again, it took some time to get the software and to get the applications user to take advantage of the processors.

I see massively parallel as going along the same lines. If I look at the supercomputer system of the future, it is going to have elements of all of those. Scalar processing is not going to go away. Massively parallel is at an infant stage now, where applications are starting to be moved into and people are starting to learn how to make use of them.

Vector processing is not going to go away either. If I look at the number of applications that are being moved to vector processors, I find a great many.

Our goal at Cray is to integrate the massively parallel, the vector processor, and the scalar processor into one total system and make it a tightly coupled system so that we can give our customers the best overall solution. From our view, we think that for the massively parallel element, we will probably have to have at least 10 times the performance over what we can deliver in our general-purpose solution. When you can take an application and move it to the massively parallel, there can be a big payback that can justify the cost of the massively parallel element.

Work that has been done by the Defense Advanced Research Projects Agency and others has pushed the technology along to the point where now it is usable. More work is going to have to be done in optical circuits. Still more work has to be done in gallium arsenide until that is truly usable.

These are all technologies that will be used in our system at Cray. It is just a matter of time before we incorporate them into the system and make the total system operative.

To conclude, Cray does believe that massively parallel systems can work. Much work remains to be done that will be a part of a Cray Research supercomputer solution in the future. We see that as a way to

move up to TFLOPS performance. I think when we can deliver TFLOPS performance, the timing will have been determined by somebody's ability to afford it. We could probably do it in 1995, although I don't know if users will have enough money to buy a TFLOPS computer system by 1995.

Massively parallel developments will be driven by technology—software and the architecture. A lot of elements are needed to make the progress, but we are committed to putting a massively parallel element on our system and to being able to deliver TFLOPS performance to our customers by the end of the decade.