Why isn't SMT Xeon scaling?

Xeon MP wheezes in

Intel has extended its SMT, "HyperThreaded" Xeon processors to large servers. But the processors don't give the dramatic performance benefits that they should, given the higher frequencies, better bandwidth and virtual processor technology.

The chips use ServerWorks GC-HE custom chipset and come in 1.4Ghz, 1.5Ghz and 1.6Ghz frequencies, with L3 cache up to 1MB.

As with the earlier SMT Xeon for wsorkstations, the chips are built to a .18 micron process, use DDR memory and the P4's 400Mhz bus. There's a significant bandwidth improvement over the PIII, too.

But does this add up to a linear scaling? Not quite.

Intel's own benchmarks show that for an extra 600Mhz, or 78 per cent increase in clock frequency, the SMT Xeon adds just 36 per cent in the SAP application benchmark. (Intel compares a 4-way 900MHz PIII Xeon box against a 4-way 1.6Ghz Xeon MP). And that's with the bigger L3 cache, and much better internal bandwidth.

So by Intel's own figures, HyperThreading isn't delivering.

We look forward to seeing recompiled applications run through the same benchmark, and word is that these do deliver the promised speed advantages. Until we do see some comparitive figures, the jury's out.

Since Windows users will need to pay a higher license fee for the virtual processors in a Xeon SMT, the price/performance advantages are beginning to look very dubious. (See SMT Xeons count double for Win2k Server licences.

The new Xeons are priced at $1,177, $1,980 and $3,682 respectively.

An additional Windows server license starts at little under the price of the 1.4Ghz Xeon MP.

Intel has vowed to base all future IA-32 server processors on SMT technology, which was formerly known as Project Jackson.

Why does Xeon MP underdeliver? Answers on a virtual postcard, please. ®