IBM is committed to deliver a "flexible server" on silicon - an interview with IBM's Michel Teyssedre

CL: Can you briefly describe IBM's vision for the High Performance Scientific field which I assume involves the RS/6000 Power3/4 processors.

MT: IBM from the very first announcement of the RS/6000 series developed a business model following a sustainable strategy. It decided to follow a unified development road map which includes both commercial and scientific customers. This provides a broader base for the RS/6000 product and avoids the instability inherent in smaller niche markets. IBM sales of the RS/6000 are roughly half to commercial customers and the rest to the technical/scientific field. In this model both customers benefit, the commercial business gains from the new developments of the ASCI programme and the High Performance Scientific end gains from affordable supercomputing made available because of high volume sales.

CL: IBM is putting a lot of effort in the ASCI programme, with the RS/6000 Power3 (nighthawk) processor used in the ASCI-Blue system with a 2.4 Teraflop/s peak performance and the promised Power4 (winterhawk), which will deliver 100 Teraflop/s in the a future ASCI system. Where is this technology leading for your normal customers?

MT: IBM is using the ASCI programme as a flagship to develop and prove new technology. The latest IBM RS/6000 Model, the nighthawk, is based on the Power3, 220MHz [4 floating point calculations per cycle] processor delivering 880Mflop/s each. This is to be used in the ASCI-White system to deliver 10 Teraflop/s next year. IBM has charted its road map with the latest Power4 1GHz chip in the pipeline which is expected to be available in volume by the end of 2001. The Power4 will have 32 processors and therefore 32Gflop/s per node. By replacing the Power3 chip with the Power4 IBM would be able to deliver 100 Teraflop/s within the existing ASCI-White architecture. IBM is also incorporating copper circuitry on silicon which is much better than the previously used aluminium. So our technology path is well known for the visible horizon. Indeed IBM promises to deliver a Petaflop/s system in 5 years time.

CL: This is very impressive but when one looks around at the major scientific/academic centres particularly in Europe, IBM is sparsely represented. Why is this?

MT: This is partially true, but IBM has 141 systems in the Top500 list, half of them commercial and half in the technical scientific field. Indeed IBM is now the number one vendor as far as the Top500 list is concerned.

CL: Yes, but still the major academic centres in Europe are overwhelmingly SGI/Cray populated, European Aerospace uses NEC SX systems and the Weather Centres straddle several vendors. Is this because of history and the time lag has not caught up with the real strengths of IBM's new systems or is it more fundamental?

MT: There is some truth in what you are saying. Also with SGI/Cray listing losses every quarter during last year one can see a business opportunity for IBM. We are talking to this sector and we are winning some of them over.

CL: Another area which traditional supercomputer users feel a bit uncomfortable with IBM, is the lack of focus and support in scientific software. I know that this may be unjust but the perception persists almost as folklore. Can you shed some light on this?

MT: Yes, IBM has thousands of scientists on its payroll and has been doing a lot of work on scientific software. We were sometimes slow to share this knowledge with our users. This has changed in the USA and will soon be changing in Europe too. IBM is creating a new group of about 20 specialists in scientific fields to focus and channel IBM's vast knowledge to the end user. This should help both the scientific community and change the outmoded perception about IBM in this area.

CL: What will the IBM RS/6000 SP product look like in the near [visible horizon] future?

MT: IBM is committed to deliver a "flexible server" on silicon. It is likely to contain two strands in the same box. One strand based on Risc technology running UNIX [closer to LINUX), and the other strand based on Intel technology running NT. It will have one Operating System and one middleware. One has to remember that the Power4 chip has a NUMA environment and binary compatibility not only within a box but also within a processor.

CL: One last question. How about the competition both in the scalable cluster systems and the shared memory parallel vector systems from Japan? After all, other vendors in the scalable domain seem to have faster chips and the shared memory parallel systems such as the NEC SX-5 are parallel, have the capability to deliver Teraflop/s, are easier to program and also have the advantage of running legacy application software without any changes. Why do you think users should choose an IBM solution?

MT: I do not like to talk about other vendors. It is up to them to argue their case. As far as I can see the vendors from the USA are moving to Intel in the near future and IBM understands this market. The Japanese shared parallel vector systems have sometime to run yet especially because of the legacy codes, but since they are not players in the US market they do not pose any threat to the overall global market position of IBM.

In my view, the production of CMOS in volume gives IBM an edge over the competition. To design a chip you need a few hundred highly skilled engineers but to move the process to fabrication and volume production you need billions of dollars.

CL: I think we explored a number of issues. Thank you for your candid answers which I hope readers would find enlightening.

[News on Advanced IT]   [Calendar]   [Analysis]   [IT in Medicine]