Test results: 1600MHz FSB
Finally, time has come for the only processor in its kind with a 1600MHz FSB. The nominal capacity of the memory controller in the Intel chipset will not be able to provide interesting results, so we'll take advantage of the flexible memory controller in NVIDIA nForce 790i Ultra. In fact, this FSB frequency limits the minimal memory frequency to 1066MHz (only in case of Intel controllers, of course), that is it's impossible to use nominal DDR2 modules here. It means that our comparison drifts from the practical field (is it justified to buy non-standard more expensive memory modules) to the purely theoretical side (which non-standard memory modules are better?) Just don't forget about DDR3 -- these are standard frequencies for such memory modules.
Well, we've got a usual situation here, similar to previous parts of our comparison: the memory read rate grows, as memory frequency is increased up to 1600MHz, but not any higher. And increased timings do not break this tendency.
The same concerns the memory write rate, only DDR3-1800 is even more useless here.
However, DDR3-1800 strikes back still in the memory read latency test: after all, absolute timings are lower in this mode.
As you may remember from our first test results of the QX9770 processor with dual-channel DDR2-800, the multi-threaded read rate reaches its maximum with two competitive threads executed on physically different cores, while the multi-threaded write rate upsprings with two competitive threads executed on cores belonging to the same physical core (sharing the same L2 Cache). Having added the NVIDIA chipset and much faster memory modules to our old testbeds, we got the following interesting results:
- Memory read rate with NVIDIA nForce 790i Ultra SLI is practically identical when two threads are executed on two physically different cores and on cores belonging to the same physical core (four-thread reading is significantly slower).
- NVIDIA nForce 790i Ultra SLI reads data from memory with software prefetch much faster, when two threads are executed on cores belonging to physically the same core (four-threaded reading is again noticeably slower).
- On the other hand, the maximum write rate with NVIDIA nForce 790i Ultra SLI is higher, when two threads work on physically different cores, four-threaded writing ranks in between.
For our purposes we'll take maximum results, obtained in slightly different conditions for multi-threaded reading and writing.
In case of the Intel chipset, advantages of using DDR3-1600 memory are apparent. Differences between various modes are not very impressive with the NVIDIA chipset, but the grand total is the same: faster memory (no faster than FSB) provides some performance gain.
All the more important to verify it in practice, and obtained results are not that promising: modes with various memory frequencies differ by 2-3%, hardly a serious impetus to buy a top memory kit.
Thus, the 'semisynthetic' group of tests corroborates our conclusion about conceptual benefits of using faster memory, maximum with DDR3-1600. But don't expect real performance advantages versus DDR3-1066. This conclusion applies not only to rare owners of the QX9770, but also to all overclockers, who seriously raise FSB frequencies to overclock CPUs.
It only remains to put together results obtained with three groups of PC configurations and correlate them with the main question of this article.
So, in case of popular Core 2 processors with 1066/1333MHz FSB, it makes sense to use dual-channel memory with much higher bandwidth than the nominal FSB can provide. If we consider DDR2-667 the reference point (as the cheapest modification in the market), fast DDR2 or DDR3 memory can bring a 6-8% performance gain in real applications. That's not bad: the difference corresponds to the CPU multiplier of x0.5. That is other things being equal, this performance gain makes the same difference as a one-step-faster CPU. And you most certainly cannot expect many-fold performance gains.
The optimal memory modules should work in pseudo-synchronous mode with FSB (their reference frequencies must match), and their absolute timings shouldn't be too high. Will such modules be expedient? Almost always not, because elite memory modules can be several times as expensive as regular modules (yielding only 6-8%). However, this conclusion will depend on the price of a whole system unit. There will be situations, when it will be the most rational way to improve your computer -- for example, if you want to buy a top processor.
Our conclusions will also hold true for overclocked CPUs. But in this case, motherboards with the most popular chipsets (Intel) won't let you use low-clocked memory. So the reference point will shift to more expensive faster memory modules. As a result, performance gains from using, for example, DDR3-1600/1800 will be significantly lower (about 2-3%), although the price difference between memory modules will become smaller.
The ATI Radeon HD 3870 graphics card has been provided by PowerColor.
Write a comment below. No registration needed!