Free performance - Hyperthreading and Turbo Boost
Hyperthreading is back!So the guts haven't changed all that much whilst the memory subsystem has, right? Yes and no. Hyperthreading is now back in Core i7, brought back from the long-forgotten Pentium 4 days.
The point of Hyperthreading is to keep the core's execution units as busy as possible by running two concurrent threads at the same time and feeding them to each core. The downsides are that on a quad-core part you need at least five threads going into the execution units for Hyperthreading to be useful and the effectiveness is negated if the threads require near-identical resources - the execution units can do more than one thing at one time but cannot duplicate resources, and causing contention for them will lead to sub-optimal performance.
Intel knows this and has increased parts of the logic to somewhat compensate for occasions when threads compete for resources. On the face of it, though, if the application can throw, say, eight threads at the processor that require disparate parts of the core's engine, then it makes a whole heap of sense, and it won't be uncommon to see gains of up-to 20 per cent when compared to non-HT situations.
Gimme Turbo Boost
Core i7's clearly been designed to add additional speed over and above Core 2, that much is clear from our dissemination thus far. We can surmise that clock-for-clock performance will be better than Core 2, and the exact amount depends upon how well the application takes advantage of the improvements listed on the previous pages.
The processor now features a Power Control Unit that moves thermal/performance control to a dedicated unit. There are real-time sensors for current, power, and temperatures. What's cool, if you'll excuse the pun, is now the ability to granularly control per-core activity, to the extent that any number of cores can be completely shut-off (C6-state, MWAIT) without impinging on others.
What this translates to, we believe, is a more-economic use of power when evaluated over a serious of applications that require different power-draw characteristics.
A clever part of the PCU is that it will apportion extra TDP headroom where and when it can. For example, should the application be threaded such that two cores are busy whilst the other two are in C6, the TDP headroom will be used to increase the core speeds to a safe limit - auto-overclocking, if you will. The dynamic overclocking is designed to keep the processor as close to the TDP as possible.
Just how high each core is pushed will depend upon the PCU's appraisal of temperatures and operating environment. It's all handled in hardware with no user intervention required, and we'll be investigating its usefulness in our performance evaluation.
