Intel Xeon benchmarketing doesn't stack up

The TPC benchmarks it gives on the page are run on four way systems using 32MB of level four cache which must throw into question whether these are real world tests.

That's all apart from the first TPC/C test, performed an an unreleased Fujitsu machine.

As we reported before, most sensible PC companies are waiting for the arrival of the Gallatin Xeon processor, which like the Pentium III "Cashcades", will have ample supply of cache to power symmetrical multiprocessing machines.

In fact, the eight way processor Pentium 4 Xeon from Big Blue uses 64MB of level four cache and just reaches 1.38 the performance of the eight way Pentium III box.

This led HP last week to savage IBM on its eight way system, forgotting of course - and as Big Blue gleefully pointed out in a memo leaked to news.com, that it has nothing comparable to offer until Gallatin comes out.

Which is just what we said a few weeks back.

The TPC/C benchmark is very processor intensive and needs 8GB to make it fly like a bird, as shown by Intel's own benchmarks. And once it's in that lovely great big memory it flies in a loop, largely untroubled by network and IO concerns.

While that might indicate that eight and four way servers using SQL databases are great given this type of configuration, most large enterprises typically use multiple applications, and anyway, Oracle is a different and more heteregeneous kettle of fish than Microsoft SQL Server.... meaning performance degradation.

The Intel tests also use hyperthreading OSes but hang on a minute, Microsoft has made its position clear on how this affects licensing. A good old four way Pentium III Xeon system doesn't suffer from this "licensing glitch">

It also seems to those close to Intel's benchmarketing plans that the SSL test it uses runs from two separate 1024 bit RSA SSL libraries for Apache, and it seems that either hard coded Pentium 4 Screaming Sindy II instructions are used, or a hyperthreading application that works is the code in question.

Good old CDC-Cray and IBM played with hyperthreading between 20-25 years ago but Intel's hyperthreading, we believe, uses two small level one caches, and that can lead to stalling on both of the CPU caches.

Intel could fix this in future designs of the Pentium 4.

Intella's P4 benchmarketing

Intel representative claim that Pentium 4 Xeons using multiprocessing are far better than Pentium IIIs, by as much as 30 per cent.

In truth, it seems that Intel needs to get its Gallatin CPUs out of the door far faster than the first quarter of next year.... µ