Skip to main content

Intel's 955X Dual-Core Chipset Better For Business Users Than NVidia's nForce4

955X Versus 945P

As with earlier generations of Intel chipsets, differences between the standard 945P and the more expensive 955X are slight. The 945P can theoretically support all Socket 775 processors, because it too supports FSB1066. But it's safe to assume that most motherboard vendors will pick 955X chipsets for their top-of-the-line motherboards along with other high-end components. This makes sense when you consider that the energy consumption of dual-core Pentium processors requires the best and most stable supporting components to head off potential timing trouble.

Differences between the two chipsets are most apparent when it comes to addressable memory. The 955X supports up to 8 GB whereas the 945 only supports half that. In addition, the 955X supports ECC-RAM, making this chipset suitable for workstation and entry-level server applications. But that makes it only more important to watch out for memory reductions from space reserved for special uses (such as PCI interfaces). Many 32 bit systems support only 3 GB of usable RAM, because the chipset doesn't support memory remapping.

The most interesting practical difference between the two chipsets is probably the Intel Memory Pipeline Technology (iMPT), currently available only on the 955X. MPT combines multiple, dynamically optimized data paths between the CPU and main memory, and employs an opportunistic instruction scheduler designed to lower latency with an improved memory retrieval algorithm to make more efficient use of memory banks as well as individual memory locations.

But after a few measurements, we observed that iMPT performance benefits are small. It's similar to differences that Intel's so-called Performance Acceleration Technology (PAT) brought to the 875P chipset, as compared to its 865 counterpart. These features may be interesting for psychological and technological reasons, but real differences between the two usually manifest only when running synthetic benchmarks (as opposed to real-world workloads).