Intel dropped the Itanium 2 into the hands of system builders a few months ago, and the first products to use the technology have now been delivered to customers. Intel's troubles have only just begun.
The mother of all chip makers says its Itanium architecture is optimised for high-performance computing, and also for business applications such as data-mining and transactional database systems. But it seems vendors that build high-end systems do not agree.
Let's look at NEC. This company markets 32-way Itanium systems alongside its SX series of supercomputers. The SX series uses NEC's own Vector processor, and the chip occupies the upper echelons of the Top500.org league table of optimised benchmark results. SX-based systems are generally used by particle physicists, weather forecasters, genetic researchers and so on. They are among the best high-performance computer systems around.
Like Itanium, the Vector architecture only works efficiently if it is running specially-optimised software. In this sense then, it looks like the Itanium is in fact designed for high-performance computing. But the current NEC chip is made up of 60 million transistors, and not one is used for cache memory. In contrast, the Itanium 2 is made up of over 220 million transistors, and the majority are used for cache memory.
The Itanium 2 has a bandwidth-to-memory of some 6.4GB/s, which is often shared between four chips in a multiprocessor system. In contrast, each SX-6 has a bandwidth-to-memory of 32GB/s. This very high bandwidth is achieved using one of the oldest tricks in the book - interleaving 16 banks of DDR DRAM.
It seems odd that two wildly different approaches may be optimised for the same workload. Perhaps this is why the chief of NEC's supercomputer division has stated that it is not possible to optimise a chip for both supercomputing and transactional database systems.
After all, high-performance compute applications are characterised as a single process that runs for many hours against a huge dataset. Transactional database systems comprise many discrete processes, each of which works for a moment on a small dataset, then waits for some I/O with one or more slow subsystems - such as a disk array or the Internet - before completing a few moments later.
Given that transactional workloads pause for I/O, it seems processor bandwidth is less of an issue in transactional systems. Perhaps this explains the success of mainframes and Sun servers for transactional workloads. Incidentally, neither Sun servers nor IBM mainframes come close to Vector-based systems in the Top500.org list.
It's also interesting that although NEC is selling a 32-way Itanium system, it is not a platform for transactional applications. Similarly, data warehouse specialist Teradata is one of Intel's biggest resellers - it buys IA-32 systems from Intel, loads its own operating system and software, and sells the result. You would have thought Teradata would have been attracted to Itanium, but it isn't. In fairness to Intel, both vendors say future versions of Itanium will become more attractive when they are delivered in a few years' time. In the meantime, few but the most committed Intel fans will be selling them as business systems.
Have your say: contact IT Week
reader comments