In the second part of our NV40 & R420 testing, we felt the CPU to be limiting performance in the lower resolutions. We'd also seen in our low latency memory round up, that having 1GB of RAM gave a performance increase in Far Cry. In general the term "CPU Bound" is being dallied around all over the place when talking about these next generation cards.

Recently, Boston sent us a rather powerful Dual 3.4GHz Xeon (Nocona) workstation on the new Tumwater chipset, with DDR-II, PCI Express and U320 SCSI. The Nocona core is basically a Prescott core with a 64-Bit Extended Memory bolt on. Considering we are running in 32-Bit and these Xeons don't have any Level 3 cache, this can be considered to be a dual Pentium 4 Prescott system hopefully offering a bit more power than our 3200+ Athlon 64. The 2GB of DDR-II memory should hopefully reduce the bottleneck of our previous DDR system and we can also reduce the AGP bottleneck by running on PCI Express. Under this system, I think we might get a better idea of what these cards are capable of.

PCI Express - The Technology

Anyone who is considering upgrading their computer in the near future will certainly have the words "PCI Express" stuck in their mind. Intel have been ready for this step forward with the 915/925 chipsets, whilst the AMD platform has been left temporarily behind until VIA and nVidia get things sorted.

The transition to PCI Express is very similar to the one we currently making from Parallel ATA to Serial ATA hard drives. PCI, ISA and AGP are built on a Parallel interface and just like P-ATA, they are reaching the limits of the technology. The solution to this is PCI Express which uses a serial interface. Unlike AGP which only offered an upgrade to the graphics port, PCI-E offers a replacement for PCI as well.

AGP uses a parallel bi-directional interface, with a maximum bandwidth of 2.1Gb/s. One of the biggest problems of a bi-directional interface, is it that when data needs to be sent in the other direction, it requires the data flow to be stopped, and data sent the other way over the same wires. This requires a certain amount of processing power to control it, and this results in data collisions and a choppy data stream. In cases like video playback, this can be noticeable.

The serial interface PCI Express uses, gets around this problem by using a separate wire for data going up as well as down. Each pair of wires is called a "lane" and has a maximum bandwidth of 2.5GB/s in both directions. In the case of a graphics card, more bandwidth is required, so a slot with 16 of these lanes is used - a 16x PCI-E slot. This brings the maximum throughput up to 40Gb/s up and 40Gb/s down (80Gb/s in total). It is this aggregation that makes PCI Express such a useful and scalable technology. Currently, In reality the bandwidth is 4Gb/s which is double that of AGP.

<< Back | 1 | | 2 | | 3 | Next >>