ARM is no fan of HyperThreading

ARM RECENTLY SAW a view of its future on TechOnline, delving into ways to improve performance of a system running multiple threads.

Simultaneous Multi-Threading, such as Chipzilla's once much-vaunted HyperThreading, absent from the Core architecture, was compared with dropping multiple cores onto a chip, to see what the pros and cons were.

Perhaps surprisingly, perhaps not, ARM (one of whose major selling points is the small die size of their solutions), state that it is considerably better to double your silicon area and stick two cores on, than it is to go for a more complex single core with SMT support, their reasoning being that a well-designed multi-core system, while bigger, will actually use less power. They claim up to 46% savings in energy over an SMT solution with four threads.

Also, moving an application to two threads on a single SMT-enabled core will increase cache-thrashing by 42%, whereas it will decrease by 37% when moving to two cores.

John Goodacre, the writer of the report, went as far as to say that SMT was only really useful as a way to rescue a modicum of performance from when you have a mismatch between the processor's speed and the rate at which you can feed it something to crunch on. In essence it was "clearly a ‘Band-Aid' measure to provide a quick fix", and had limited longevity and applicability.

While no doubt this is a benchmarketing exercise to promote ARM's tiny, but non SMT-capable, ARM11 MPcore designs, Intel can't disagree too hard, or we'd have seen some form of HyperThreading in Cointreau! But the rumours still circulate about its possible return in the next generation microarchitecture. µ

L'INQ
Tech Online You have to register