IBM, Moore's Law and the POWER 5 chip
The benchmark to end all benchmarks
IBM's recent pSeries benchmark ought to raise a good number of eyebrows. Unix server benchmarks had previously been a game of leapfrog between IBM and HP, with IBM looking increasingly strong in the last round (in 2003) when its p690 server with 32 processors demonstrated a slightly higher transaction rate than HP's 64 processor Itanium 2 server. The latest benchmark from IBM leaves both previous benchmarks in dust. In the TPC-C benchmark, IBM demonstrated over 3m transactions per minute (tpmC) almost three times more than the previous highest result of 1.2tpmC, posting a 37 per cent price performance advantage in the process.
Never in the history of the benchmarking of servers (or databases) has any company demonstrated such a dramatic lead over the competition. When one vendor leapfrogged another by ten per cent or so (as regularly happened) you could argue that the difference was irrelevant. A benchmark is, after all, only a benchmark and technology just gets faster, doesn't it? But a difference of this magnitude cannot be dismissed in that way. IBM has clearly achieved something dramatic with its technology. If you analyze this benchmark result in any depth then you quickly conclude that there are several contributory IBM technologies that helped to achieve the result. However, here let us focus on what is, perhaps, the major one: the POWER 5 chip.
At a recent IBM analyst conference, Bernie Meyerson, who heads up IBM's semiconductor R&D development, gave a good explanation of what was happening at the semiconductor level. Moore's Law (which proposed that computing power would double every 18 months or so) was obeyed for decades in large part by miniaturisation - putting a lot more transistors onto the silicon with each new generation of chip. However, as he explained, miniaturization is a little more complicated than shrinking components to a smaller size. There are, in fact, 15 dimensions or characteristics that determine the behavior of the micro-circuitry that need to be kept in step.
Well, that was not too much of an obstacle until things got really really small and you had to deal with layers of material which were no more than 5 atoms thick. When you get down to that level, as has now happened, life gets complicated. The major problem that emerges is voltage leaks, which cause the chip to heat up considerably and ultimately put a block on progress. This is not the only problem, but it is the most difficult one to address. The upshot is that further miniaturization is not achievable without considerable innovation, particularly a different approach to power consumption.
Bernie Meyerson's message was simply this:
"Moore's Law is no longer guaranteed."
So how come, at this point in time, is IBM able to produce a blistering benchmark with its pSeries and POWER 5 chip?
Meyerson and his team saw the problem coming quite a few years ago and prepared for it in advance in the direction that he drove the PowerPC chip technology. I'll skip the full technical details, but to summarise:
Meyerson's team implemented innovations in on-chip design to cater for power management and address other miniaturization challenges. On top of that the design of the chip took a "holistic route" with functionality being added to better enable other IBM technology initiatives - such as the development of its hypervisor technology and the introduction of logical partitioning.
The consequence is that IBM is currently able to deliver in line with the expectations of Moore's Law and will probably be able to do so for the next two generations of the PowerPC chip. In fact, in terms of this benchmark, IBM is delivering beyond the expectations of Moore's Law. While chip performance may have doubled every 18 months in the past, system performance has not and IBM has now delivered almost a tripling of system performance (and a 37 per cent price performance improvement). That is what the benchmark indicates.
And IBM is doing this at a time when its primary competitors are "feeling the heat". IBM has demonstrated that it can overcome the problem, for the moment at least. It will be interesting to see how Hewlett-Packard, Dell, Sun and Intel respond to this.
Copyright © 2004, IT-Analysis.com
Sponsored: Hyper-scale data management