The SPP Memory Controller

Most of the engineering efforts of the nForce 4 Intel platform is the implementation of the memory controller. We have seen fairly small differences between most Athlon 64 platforms because the controller is integrated into the processor negating much of the difference that is seen between different motherboards outside of the subsystem. This is not the case on the Intel side and the performance differences seen ahead will be largely due to the implementation of the SPP. The implementation of the memory controller seen on the nForce 4 for Intel is interesting because it takes a different design approach than what has been seen with Intel boards as well as the older nForce 2. NVIDIA highlights several key points with their memory controller implementation:

While a dual channel memory architecture is not really anything new, NVIDIA introduces a bit of a refinement. Memory banks do not need to be populated symmetrically. This means that while two DIMM slots must be filled in order for dual channel mode to work properly, it means that the DIMM modules do not need to be identical. It also means that adding a third stick or fourth stick of memory will not push it back into single channel mode.

Official high-speed memory support is another addition. Intel's 925XE officially supports up to DDR2 533Mhz but the nForce 4 goes one step higher to DDR2 667. There is also an emphasis on low latency DDR2 modules. NVIDIA has been working together with Corsair to validate high performance memory for the nForce 4 for Intel with timings of 4-4-4-4 and 3-2-2-2. There seems to also be an implication that the nForce 4 is the only current Pentium 4 core logic chipset that allows for 1T timing - neither the Gigabyte 925x board or the ABIT 925XE based boards that we have allow for 1T. We have seen the effects of 1T versus 2T timing on the Athlon 64 in a recent memory review and readers are encouraged to take a look at the performance differences seen. The advantages of 1T timing is explained from this following quote from NVIDIA -

1T address timing is achieved when the address and commands are placed on the address bus by the memory controller and latched in by the DRAM devices, all in one clock cycle. The other alternative is 2T address timing, where the memory controller places the address and commands on the address bus in one clock cycle and the DRAM devices latch them in the next clock cycle. 2T address timing is used to ensure adequate setup and hold times for the address and command. This is necessary when the address bus is heavily loaded, which usually occurs when the address bus is shared by multiple DIMMs.

In essence, 1T timing reduces memory latency and the design of the nForce 4 memory controller allows for this whereas the current Intel chipsets do not. To take advantage of 1T timing, NVIDIA also uses a shorter burst length - with a 2T system this will increase latency further.

Two further pieces make up the memory controller. The first is DASP, NVIDIA's Dynamic Adaptive Speculative Preprocessor which is a fancy name for a fancy technique of treating system memory like a L3 cache by predicting what a thread or process will likely need. By keeping the most used data in the memory controller, latency is once again reduced and system performance is claimed to increase by 2-4%.

The final piece of the equation is dubbed QuickSync. Back in the days of the Athlon XP, it was always suggested that the FSB and Memory were run at a 1:1 ratio because attempting to bridge an asynchronous processor bus and a memory bus is very tricky. The nForce 4 SLI for Intel solves this problem and maximizing both the processor bus and the memory bus speeds will increase system performance regardless of the other setting.

Media Shield

Just like the way that everyone has their own name for their storage technology, NVIDIA is no different and has gone with Media Shield as the all encompassing label for some of the value added features including RAID, Disk Alert (a visual representation of which exact disk has failed), and a revamped, more user friendly interface.