I/O Testing and the NVIDIA Press Kit
Just a quick note about the I/O performance of motherboards and chipsets on those boards. From now on at HEXUS, we'll seek to cover performance of the peripheral I/O busses provided by a motherboard. It's not enough to see 8 USB2.0 ports on the spec sheet. Implementation differences mean that those eight ports on one motherboard may enjoy lower CPU utilisation and higher measured bandwidth than on another. Worth testing.So we'll endeavour to test network I/O, ATA/SATA, USB and FireWire performance using a number of methods. While all of those methods won't appear in this article - testing hardware needed for attached disk storage had annoyingly gone AWOL at the time of writing, and development of network I/O benchmarks still being in progress - they should all show up in future motherboard and chipset analysis, all things going well.
What I will say, with regards to some quick USB2.0 testing, is that it appears remarkably low in terms of CPU utilisation, more so than recent boards based on nForce2 Ultra 400, SiS655FX and VIA K8T800. Testing of a retail board will hopefully show that to be the case on other implementations.
Testing with NVIDIA's Press Kit
NVIDIA's full press kit for the nForce3 250Gb arrived shortly after benchmarking the qualification sample. With unlocked Athlon 64 Model 3400+ (able to run at 1000MHz HT clock), Gainward GeForce FX5950 Ultra, 512MB of Corsair XMS3200 and RAID0 36.6GB Western Digital Raptors, it was sufficiently different from our usual test setups to negate its appearance in the benchmark graphs.However being a 250Gb part and arriving with RAID0 Raptors, it was useful in testing the Gigabit Ethernet implementation and hardware firewall, plus the new native SATA150 controller in RAID form. HDTach 2.61, rather than the retarded 2.7 version, lets us measure the performance of the Raptors on the new SATA controller.
We're looking for high burst speeds mainly, measuring its ability to transfer as much data as possible over the controller interface. A flat graph with few peaks and troughs (although hard with RAID0 due to its striped nature) is a bonus too.
Up and down like Zebedee on crack, the graph is nonetheless representative of good performance from a RAID0 stripe set of Raptors, based on results you'd get on other controllers. With the qualification sample, it was hard to quantify disk performance of the new SATA controller, partly due to lack of suitable disks, partly down to the early nature of the board itself. With NVIDIA's full-on press kit, it's easier to evaluate. Initial performance seems strong, with ~185MB/sec burst speed (notice > 133MB/sec), a useful average speed rating over and above a single disk, along with low random access time.
CPU utilisation isn't the lowest you'll ever see, but the nature of blocking I/O requests means that as long as CPU usage isn't too high, you rarely notice the difference between 5% and 25% CPU utilisation.
What about the network part of things with 250Gb?
