PCI Express Test
The problem shows up more clearly in the PCI Express bandwidth test. The point of this test is in sending a certain amount of data from the system to the graphics memory and back again, the size of the transferred block varying from 64KB to 4MB. The utility makes use of the memory controller as well as the graphics bus and the bus that connects the CPU and the chipset (in this case it is the HyperTransport link). Here are the results the test produced on the nForce4 SLI and nForce4 SLI X16 platforms:
The mainboards do not differ much when the data block size is the smallest, although the ASUS A8N-SLI Premium is ahead of the ASUS A8N32-SLI Deluxe even in this case. The negative effect of the 8-bit HyperTransport channels between the CPU and the nForce4 SLI X16 North Bridge is concealed by various time latencies. At the biggest data block size this loss is minimal, so the architectural deficiency of NVIDIA’s new chipset shows up in full. The ordinary nForce4 SLI connected to the central processor via two 16-bit HyperTransport channels is 20% faster than the nForce4 SLI X16!
It’s the same when the data is transferred in the opposite direction, i.e. from the graphics card’s onboard memory into the system memory, except that the mainboards have nearly the same performance on smaller data blocks.
Theoretical Benchmarks Summary
Our theoretical tests have showed that the new chipset from NVIDIA and the A8N32-SLI Deluxe mainboard based on it do not have any advantages over the older nForce4 SLI-based platform. In some cases, when there are large amounts of data traveling from the chipset to the CPU and back, the new mainboard is even slower than the older one because its HyperTransport bus between the CPU and chipset is two times narrower.
We want to remind you that the ordinary nForce4 SLI uses two 16-bit channels with a bandwidth of 8GB/s to either side while the nForce4 SLI X16 has two 8-bit, 4GB/s channels and the total bandwidth is 8GB/s against the nForce4 SLI’s 16GB/s. This design difference didn’t show up much in the theoretical tests, though. Will this be so in gaming applications, too?