Core i7-975 Extreme Details

The Core i7 vital stats you already know apply here, just as they have since the architecture launched last year. Manufactured on Intel’s now-mature 45nm process, a single Core i7 die populates 263 square millimeters.

Natively quad-core in that the processor doesn’t consist of two dual-core die on a single multi-chip module (like Core 2 Quad), Core i7 also includes Hyper-Threading technology. The result is a micro-architecture with four physical execution cores able to concurrently work on eight threads. As a result of the efforts expended by software developers to better-optimize relevant code for parallelism, this results in a performance win more often now than it did back when Hyper-Threading first emerged in the Pentium 4 days.
Cache sizes remain the same (32 KB L1 I/32 KB L1 D and 256 KB L2 per core, plus 8 MB shared L3), and the integrated triple-channel memory controller is still officially limited to DDR3-1066. But of course, as we’ve discovered, retail CPUs support the multipliers necessary to reach as far as DDR3-2133. We have heard from one memory vendor that the controller itself has been improved, but without any additional information from Intel regarding how it might have been tweaked, we can’t confirm those rumors at this time. We can say that DDR3-2133 is now within reach, though it takes some serious tuning to stabilize at that data rate.

As with the i7-965 Extreme, the 975 boasts a 6.4 GT/s QPI link, while the i7-950 employs a 4.8 GT/s link. Of course, if you’re running a retail processor (and not an engineering sample, like the one used in our original Core i7 launch coverage), you should be able to manually tune QPI speed up to 6.4 GT/s in your motherboard’s BIOS.

We asked Intel about the i7-975's Turbo bin configuration and were told that it is exactly the same as the i7-965 before it. That is to say, when 1, 2, 3, or 4 cores are active, you get 2, 1, 1, and 1 available bin (a bin being 133 MHz). Curious as to how much time our 975 Extreme would spend at 3.6 GHz, we ran a single thread of Prime95 to tax an individual core. Interestingly enough, you spend a lot of time waiting for that 27x multiplier to kick in (up from 25x), and it doesn't last very long. You see, there's always something else going on in the background, and if there isn't a significant load being applied to at least one thread, SpeedStep is throttling you back the other way. Expect most of your load time to be spent at 3.46 GHz with Turbo mode enabled. Otherwise, turn the feature off completely and overclock manually.
As a result, we have to wonder how much benefit upcoming architectures will see from Turbo with a single core active.