Features:

LINPACK RESULTS REFUEL IBM/INTEL CHIP DEBATE
By Christopher Lazou, HiPerCom Consultants

The Intel Itanium 2 (1.5GHz Madison) chip was only announced 30th June, and already some vendors have announced their new Madison systems, tailored using specially designed chipsets. Examples of these include the 64 CPUs HP Superdome, the 64 CPUs SGI Altix and the 32 CPUs NEC TX7/i9510.

The new updated LINPACK-HPC benchmark released by Jack Dongarra, 15th July, naturally includes results from some of these new systems. For example, the 64 CPUs HP Superdome with Rmax = 335.45Gflop/s is on the list and is currently the fastest configuration in a single system, delivering 87.36% efficiency. A new entry in the list is the 32 CPUs NEC TX7/i9510 with its Rmax = 171.61Gflop/s is also currently the worlds fastest in its 32 CPUs class, delivering 89.38%, efficiency. Note that no LINPACK results are available on the 15th July list for the announced 64 CPUs SGI Altix, system.

These entries are of course in the "a look at parallel computing section" where massively parallel computers are compared. Interestingly, the 32 CPUs IBM eServer pSeries 690 Turbo (1.7GHz Power4+) chip, with Rmax =143.3Gflop/s, delivers only 65.85%, efficiency. Thus the NEC TX7 Madison is 20% faster than the IBM Power4+, which has a faster clock frequency. This is clear evidence that, a computer system is not just a chip with its performance depending solely on clock frequency. Other elements in the chipset environment, memory bandwidth, interconnect and so on surrounding the processor chip, also make a difference.

With low growth in the computer server market the chip competition is becoming ferociously acute. In the 64-bit market, (excluding the low volume NEC vector supercomputers) the big players are IBM with the Power series chip and Intel with the Itanium 2, product line. The SGI MIPS and the Sun Microsystems SPARC chips are minor players and in my view, unlikely to make a significant comeback. Similarly the AMD Opteron chip, the new kid on the block, it has a long way to prove its capabilities against Xeon-based systems from Dell and IBM. Having said this, Fujitsu is also producing a SPARC chip, the PrimePower HPC2xxx targeting the SMP cluster market and has the potential of a major player. So, with all these chips around and the Madison chip already integrated into new computer systems, which at least on the LINPACK benchmark are outperforming the IBM Power4+, does this spell trouble for IBM?

At this point a word of caution about benchmarks is in order. Most of you have seen marketing slides with selective benchmark results, which tend to subvert reality, but purporting to come from procurements. These results often do not compare like-for-like systems and because there is only one winner, confidentiality prevents us to know whether the choice of system was made using the best cost/performance integral, or for some other more nebulous reason. With all its acknowledged limitations, the LINPACK benchmark, at least requires vendors to submit their computer system to the same strict measuring conditions, so that a comparison can be made with some degree of confidence.

According to a reliable source at a recent Palisades briefing related to chips, IBM asserted that in the chip business, the real competition takes place at the fabrication level. Building new fabrication facilities costs a billion dollars or more and you need to sell all the output from the fabrication plant to amortise costs. You only build another fabrication plant when you are sure demand is picking up and you can sell the chip output from it. Until recently IBM was bullish about withstanding pressure from Intel entering the 64-bit market. Karl Freund, a VP in the pSeries server group, claimed that IBM is outperforming Intel in the 64-bit market. If this turns out to be true and persist in the future, it would have dire consequences for HP, which put all its eggs in the Intel Itanium basket. With the Superdome delivering 87.36% efficiency, this is however, unlikely.

Karl made this assertion before the Madison was announced and the Madison based new crop of computer servers, arrived on the market. He also claimed that the pSeries Power chip has the edge on the Intel Itanium line because the Itanium is a "hotter" chip and in order to get higher performance in the future, the Intel Itanium chip will have to become water cooled, rather than remain air cooled as at present. That would increase costs and would also be very inconvenient especially for blade servers. As is well known extraction of dissipated heat is a deadly scourge of chip production. Karl claimed that the IBM Power chip is one generation cooler, so IBM will only be forced, into water-cooling, a year or two after HP treads that path, using the Intel Itanium successor products. For an outsider, an alternative question immediately comes to mind. Will the Intel juggernaut sweep the 64-bit chip market in its wake?

I suspected the IBM claims are bravado and paint a somewhat rosy scenario for the IBM pSeries. So when the opportunity arose, I asked Peter Ungaro from IBM and Richard Dracott from Intel who were fielding answers, to comment and clarify the situation. The answer from Peter Ungaro was very diplomatic, saying that IBM is also producing and marketing Xeon-based Intel systems. He also pointed out that 77% of IBM held patents have gone on the P4+ and P5 single system image developments, which means IBM should be in good shape to deliver competitive HPC computers. They both claimed that their intention is for new generation chips to stay air-cooled as long as possible. I had to be satisfied with their answer without asking how long is a string!

As to what the future holds for IBM as a capacity 64-bit chip vendor, with Intel promising that the successor Montecito chip is to have three times more performance, compared to the Madison, these are interesting times and time is a great leveller of claims and opinion. But, perhaps both vendors should recall the Dahrendorf dictum: "History proceeds by changing the subject", and not despair.

The recent (July 16) deliberations by the U.S. House of representatives show that capability computing is attracting "leadership" attention, with many potential changes in store, but comment on this issue has to wait for another article.