What is Write Combining?

[corkyg]

In Radeon Advanced Display settings, in the Troubleshooting Tab is a slider to vary video acceleration. Below it is a checkbox for "write Combining." What is that, and how does it effect things?

 

[corkyg]

Well - it's worth one bump - seems like all my simple questions are too technical for this forum. Too bad ATI does not define what all their settings do either in their Radeon manual or on-line help.

 

[Jeff7]

Based on this, it looks like a performance tweak.
Looks like it's a tweak from several years ago; don't know that it would do anything these days. Only way to tell might be to reboot, run a benchmark, then enable write combining, reboot, re-run benchmark, compare results.

 

[corkyg]

Thanks, Jeff7 - that all makes a lot of sense. My conclusion is that with an AMD processor, it is really irrelevant anyway. I have tried it both ways and can detect no difference in performance.

 

[BFG10K]

Write combining is the process of combining smaller data chunks into larger, burstable transactions to improve performance and traffic throughput. It was really popular in the days of 3D software rendering where games relied on writing out huge quantities of raw pixels across PCI/AGP because write combining would speed things up immensely.

Nowadays I think it still helps with PCI and AGP traffic but it probably doesn't have the same impact for games like it did before. It should still help a lot for 2D operations though.

 

[Peter]

Write combining is important when data outgoing from the CPU are heading for an I/O device, like a VGA card. This is uncached area, meaning that the write data would go out one by one, as single 8-. 16- or 32-bit data. This makes very poor use of the CPU and PCI (or AGP, same thing) busses, since the best performance is only achieved when issuing larger chunks of data in one so-called burst, which is 8 chunks of 64-bit data.

Now you can teach a CPU which I/O allows write data to be collected into burst chunks - before being presented on the CPU bus - without affecting operation of the I/O device in question. VGA cards' linear frame buffers are a prime candidate for this - it's actual RAM, so that reordering and collecting write data does not have any side effects, and there are lots of write data going out.

This is PARTICULARLY effective with modern CPUs that have lots of buffers, deep FIFOs, and fast front side busses.