"We would ship the next-generation super computer by end 2007-08"

Silicon Graphics as a name has become almost synonymous with High Performance
Computing, especially for the entertainment industry. Almost all leading global
studios are today working closely with SGI-especially with the Digital
Intermediate services and Prism software released recently in Hollywood
(reported in the Jan 15 issue of Dataquest). Dr Eng Lim Goh, senior vice
president and chief technology officer, Silicon Graphics Systems, now speaks to
Dataquest about SGI's strategy to address the growing scientific and technical
computing market with its new supercomputer and about its open source
initiatives.

Silicon Graphics (SGI) has been talking for sometime about its Project
Ultraviolet that envisions a new multi-paradigm computing architecture. Can you
elaborate on this?

Project Ultraviolet is the code name given to Silicon Graphic's efforts to
design the next generation supercomputer planned for shipment by the end of
2007-08. Using a new concept called "multi-paradigm computing", these
supercomputers would unite previously disparate computing architectures with SGI's
scalable shared-memory technology. Our long-term vision is to improve
productivity by creating the first supercomputers capable of supporting and
combining different computational approaches, providing optimal performance
specifically for technical applications, regardless of programming model or
computational balance.

How this architecture can reduce the gap between actual applications and
peak compute performances?

Today, scientists and engineers make machine purchase decisions that commit
them to adapting or writing their applications, with life spans of ten years or
more, to match their chosen architecture. Moreover, the choice among clusters,
scalar shared-memory, or vector architectures often restricts the types and
sizes of data models that the adapted application can handle, due to inherent
machine limitations. This is increasingly constraining progress, particularly
with growing interests in the advanced computational techniques of stochastic,
ensemble, multi-physics, multi-scale simulations and multi-disciplinary design
optimization.

Multi-paradigm computing will change that by efficiently supporting the
different paradigms at a fundamental level. Therefore Project Ultraviolet is
based only on performance and SGI hopes it would help scientists in the domains
of manufacturing, life sciences and earth sciences especially to become more
productive by letting them focus on science, and not computer science.

Can you explain how multi-paradigm computing architecture encompasses both
scalar and vector mode?

Our team came up with an analysis showing that even for scalar systems such
as SGI's, and for vector systems, there will be significant fall-offs in
efficiency going forward into the next decade. Scalar systems are extremely
efficient when performing arithmetic (including floating-point and particularly
integer) operations. But they are less efficient handling memory accesses. The
converse is true for vector systems.

Dr Eng Lim Goh SGI's open source initiative is not restricted to XFS and OpenVault; it has been a long time contributor to the Linux community

If we looked at the way a single application runs, there is going to be a
portion of it that is best run on a scalar processor-based machine, such as
those from SGI, HP, IBM and Sun. But there will be also another portion best run
on a vector machine, such as those from Cray and NEC, and no one paradigm has a
monopoly on across-the-board efficiency.

For our customers in the next decade, we revisited the benefits of scalar vs
vector. Instead of deciding which approach fits the broadest set of needs, we
ended up saying why pick just one or the other? All in all, it is clear that no
single paradigm platform will have a monopoly on efficiency. So we decided to
come up with something new-a system that offers the best of more than one
world: a tightly-integrated multi-paradigm machine.

This architecture we are moving towards will tightly integrate, into two
chips, all four paradigms. Firstly, the scalar paradigm, using one of the
industry's fastest commodity 64-bit scalar processor-cores that is contained
in an Intel socket. This will form the foundation of the machine. For the
remaining three new paradigms, they will be built into our separate Hub chip
that is directly attached to the Intel socket. They are the vector, PIM
(processor in memory), and application-specific (allowing for directly
attachable FPGAs and user ASICs etc) paradigms. During run-time, as the needs of
the application change, the different paradigms best able to perform the task at
that moment may be switched on. Thus increasing application performance and
machine efficiency.

SGI's interest in open source is reflected by its transfer of some of
its key technologies to the Linux open source community. Can you elaborate on
this?

SGI plans to help scale Linux to run enterprise-class applications by delivering
technologies from its core competencies in visualization and high-performance
computing. XFS is one of several of these key open source contributions. SGI
plans to help Linux mature into a world-class, robust, reliable, feature rich,
widely used, and enterprise-ready operating system. SGI is contributing the XFS
filesystem to the open source Linux community using the copyleft GPL license.
XFS is one of SGI's core competencies in high-performance computing. It
combines advanced journaling technology with full 64-bit addressing and scalable
structures and algorithms. XFS uses efficient table structures for
fast searches and rapid space allocation. XFS continues to deliver rapid
response times, even for directories with tens of thousands of entries.

We also released the standards-based OpenVault source code into the open
source software community to provide developers with a removable media
management software solution for SANs. SGI hopes that by using open source
resource sharing, storage management vendors can reduce application development
costs and increase the number of time-to-market SAN solutions available to end
users. As the storage industry moves toward SANs, OpenVault open media manager's
heterogeneous technology is designed to allow removable media and robotic
libraries to be available-through a standard API-to all hosts, requiring
minimal effort to optimize these expensive resources.

However, it would be wrong to assume that SGI's open source initiative is
restricted to only XFS and OpenVault. SGI has been a long-time contributor to
the Linux community. Over the past five years, we have offered key graphics
technologies to the open source community, including Open Inventor, an industry
leading visualization scene graph, an OpenGL sample implementation, and the key
components for the OpenML standard.

Rajneesh De