logo
EnterTheGrid - Primeur Live!

EnterTheGrid - Primeur is the premier Grid and Supercomputing information source in the world. With Primeur Live! it brings you Live reports from Europe's main Supercomputing and Grid events

>Primeur Magazine
>PrimeurLive!
>EnterTheGrid
>Analysis
>Backissues
>Calendar
>Subscribe
>Advertise
>Contact
Issue 27 June 2003
>Start
>A new design for supercomputers?
>Focus
>GRIA takes Grid computing into the real world
>It is hard work to keep up with people expecting us to follow Moore's law
>TOP500 supercomputing
>Off-the-shelf supercomputing is a dead end
>Interdependence of architecture and software for effective terascale computing
>Building a PetaFlops class machine for large scale system design experience and biomolecular simulation
>Exploring the benefits of FPGA-processor technology for genome analysis at Acconovis
>Twenty years experience at NAL with software for HPC in aerospace science and engineering
>Software for large-scale computing: it is scalability that matters!
>Can SuperData Centres be secured?
>Complexity of data in the passenger services systems of the DB AG
>Billing of million customers at German Telekom
>The Grid
>Taming huge data volumes
>Company news
>Rapidly evolving microprocessor technology turns throughput computing into alternative for HPC
>Dell introduces 64-Bit server for high-performance computing market
>Efficient network-storage, TCP processing and processor development under the loop at Intel
>AMD Opteron processor answer to tough challenges in high performance computing
AMD Opteron processor answer to tough challenges in high performance computing
Heidelberg 27 June 2003 Michael Goddard from AMD introduced the new AMD Opteron processor to the attendees of the ISC 2003 Conference. As detailed as he was on the AMD Opteron processor features, as little he would expand on performance and software infrastructure. Let us have a look though at what Mr. Goddard had to say about the six challenges that advanced multiprocessor systems have to deal with.
Advertisement
Visit our sponsors
Advertisement
Dolphin's SCI interconnect features the lowest latency and wire speed

The speaker stated that x86 is the most widely installed instruction set in the world but it is not basically relevant to CPU performance. On the contrary, what is really important are backward compatibility to x86-32, the cost per processing node, memory bandwidth, addressable memory, the I/O infrastructure, and watt density.

Mr. Goddard explained that with an increased system memory data intensive applications are installed with strides and block sizes that can cause cache thrashing. Making the cache larger is not cost effective, according to the speaker, so performance is limited by the size of the on-chip cache and/or the memory bandwidth. One can enhance the memory bandwidth and latency with an on-chip memory controller which scales with SMP.

Furthermore, due to cost/performance and I/O constraints, x86-32 clusters are limited to two processors per node. This is putting additional stress on the SMP cluster interconnect. The speaker believes that it is useful to bring 4 and 8 processor SMP systems closer in cost/performance to 2 processor systems. In order to improve performance and decrease premium without breaking the x86 commodity economics, one has to use the same processor architecture on the desktop.

There is also an enormous investment in x86-32 for all market segments. In many applications, the porting code is not an option, according to Mr. Goddard. Therefore, AMD aimed at providing a solution that was not only 100% backwards compatible, but designed to run x86-32 code faster than any existing x86-32 architecture available. The company has offered a gradual and controlled migration path for porting to 64-bits, allowing simultaneous 32-bit and 64-bit application execution.

The AMD Opteron processor's performance-enhancing features include:

  • Optimized 64-bit core with uncompromising support for 32-bit applications
  • High-bandwidth integrated memory controller scaling with processor frequency and number of processors
  • Highly scalable HyperTransport technology-based system bus enabling glueless multi-processing
  • 1MB L2 cache
Mr. Goddard ended his talk by saying that the AMD Opteron processor is designed to maintain compatibility by leveraging the existing infrastructure and running existing 32-bit x86 applications natively. It also allows customers to migrate according to their schedule with a low learning curve for users and support staff.
Advertisement
Visit our sponsors
Advertisement
Leslie Versweyveld

EnterTheGrid - Primeur

James Stewartstraat 248

1325 JN Almere

The Netherlands

http://EnterTheGrid.com

mailto:primeur@hoise.com

© EnterTheGrid - Primeur Live!