In addition to the above described memory controller optimizations, nForce4 SLI (Intel Edition) also took advantage of the good old NVIDIA?s trump: the DASP unit (Dynamic Adaptive Speculative Preprocessor). DASP is responsible for prefetching from the memory the data that is predicted to be requested by the CPU any time soon. This allows reducing the CPU idling time while it is waiting for the data to arrive for processing. In fact, all contemporary CPUs have their own data prefetch mechanisms, which allow them to move data into the cache memory. However, these algorithms are not 100% efficient. This is exactly when DASP unit helps a lot by making the corresponding memory pages available for the CPU upon request.
The DASP unit of the new NVIDIA nForce4 SLI (Intel Edition) chipset is marked as version 3.0. In fact there have been pretty significant improvements made to it since the times of the nForce2. First of all, I would like to draw your attention to enhanced prefetching algorithms, which precision was improved by extra statistical analysis of memory requests submitted by contemporary applications. Also NVIDIA engineers took into account the Hyper Threading technology supported by Intel processors as well as the upcoming dual-core architecture of the Intel CPUs. So, the new DASP 3.0 can recognize different requests sent to the memory by each of the cores (physical or virtual) in order to improve the precision of the data prefetch algorithms.
As we see, the memory controller of NVIDIA nForce4 SLI (Intel Edition) chipset looks very attractive from the theoretical point of view. However, we will be able to declare our final verdict about its efficiency only after a series of practical tests. That is why we carried out some tests of the memory subsystem, which should allow us to evaluate objectively the efficiency of the new NVIDIA chipset memory controller.
The results obtained in SiSoft Sandra 2005 are quite ambiguous. When we used a Pentium 4 660 with an 800MHz bus, the NVIDIA nForce4 SLI (Intel Edition) chipset showed its best result with DDR2-667 SDRAM and with DDR2-533 SDRAM it falls a little bit behind i955X. However, when we performed the same test with the Intel Pentium 4 Extreme Edition processor working at 3.73GHz and supporting 1066MHz bus, the situation changed. NVIDIA nForce4 SLI (Intel Edition) chipset remained an indisputable leader with DDR2-667 SDRAM, but lost to all its rivals once we switched to slower DDR2-533 SDRAM.
I believe ScienceMark 2.0 benchmark could raise the curtain of mystery above these results. As we see, the bandwidth of the bus between the CPU and the memory is really high by nForce4 SLI (Intel Edition). However, once we start testing the latencies, NVIDIA chipset turns out unable to compete with the Intel solutions, especially in case the bus frequency is 1066MHz. At the same time, however, we should give due credit to extremely low latencies of the nForce4 SLI (Intel Edition) memory subsystem when we have 800MHz FSB and DDR2-667 SDRAM installed.
The obtained results allow us to make quite unexpected conclusions. The memory controller implemented in the SPP of the new NVIDIA nForce4 SLI (Intel Edition) chipset is truly a success, as the benchmarks prove. However, it is remarkable that provided the processor bus bandwidth and the memory bus bandwidth match, the NVIDIA chipset starts falling behind its Intel rivals. In particular, we see this happening in case of 1066MHz bus and DDR2-533 SDRAM. It looks like the CPU loads the entire memory bus bandwidth thus blocking the work of such intellectual algorithms of the chipset as DASP 3.0, for instance.
