Dr. Bennett sees multi-core everywhere. The dual-core Intel processor volume is exiting in 2006. There are 17 multi-core CPU projects with already 4 in production.

The speaker pointed out the relentless pace of change. In 1993 the Paragon XP/S 140 was nr. 1 in the TOP500 list with 140 GFlops and 1840 sockets. The machine filled a room. One year later in 1994, the nr. 1 position was taken by a COTS personal cluster with 192 GFlops and 8 sockets and only filling a suitcase.

Dr. Bennett sketched the example of a man who gets stuck in a traffic jam on his way from the airport. On top of this his car is hit from behind in a typical tail-collision. He wants to call his Polish colleague and tell him he will be late for their appointment. He has to make the call handsfree but he cannot accurately pronounce the name of his Polish friend for the autocall system to make sense of it. In addition his Polish knowledge is not sufficiently up-to-date to make himself understandable to his friend. What this man needs is a phone with a built-in automatic calling and translation function. Just to make obvious that people expect to have available computational power at their fingertips.

Dr. Bennett also showed a chart of the growing number of clusters figuring in the TOP500. Clusters are cheap and built out of commodity components. Their rise starts in 1995. There is a spectacular growth in 2005. We see an acceleration of their appearance after eight years.

The speaker went on about HPC codes in industry. In practice scaling is limited with 24% scaling to only 1 processor and only 59% scaling to 16-way or less. 71% of the codes in the study were developed in-house by the ISV. There are 12% in university and 5% in national labs. 78% of the codes have had a "light" refresh within the last year, 5% not for over 5 years. There have been no major rewrites.

So what needs to be done, according to Dr. Bennett. He stated that multi-core means multi-threads and not just for HPC workloads. The end of the decade will see a minimum of 32 threads in a typical 4S server. The majority will not be HPC workloads. Investments in scale-out are fundamental. With the hardware no longer the main differentiator, it will be algorithms and code.

Here the speaker cited his former English teacher who taught him that an infinity number of monkeys and an infinite number of typewriters will eventually generate the entire works of Shakespeare. However, an

inifinite number of programmers and an infinite number of nodes will not generate the entire world's software.

HPC is a leading indicator for the rest of the technical industry. Dr. Bennett stressed that we should invest now in algorithms and software. Hardware is becoming a commodity. He concluded that Intel in not just the world's most prolific supplier of HPC buildng blocks and he invited the audience to come and talk to him and his colleagues about software.