Ganged versus Unganged Memory
In the BIOS there's a setting to enable or disable an unganged memory controller. Phenom has two memory controllers inside and usually these will be “ganged” together, meaning they both work together in dual-channel, with data from all the cores sitting in one queue while they wait for a memory access cycle to become available.
However, the alternative is to enable unganged mode which means the two controllers work independently, at potentially different clocks, and each gets a DIMM (or two) to use in its own channel. You lose overall bandwidth, but single-channel performance is usually more efficient and the cores can potentially get access to memory sooner now that there are two of them.
It’s been reported that the ganged mode usually works better for single-core performance, but unganged works better for multi-core. While we ran all the tests using the BIOS default (ganged memory), but it’s down to the motherboard manufacturer to decide what option it sets as standard or even if it offers the option in its BIOS at all.
Let's have a look at how this option affects performance...
-
Ganged Memory
-
Unganged Memory
0
2500
5000
7500
10000
12500
MB/s
In the results you can clearly see that single channel performance is quite a bit more efficient than dual channel in unganged versus ganged. Unfortunately the current 4.2 edition of Everest only loads a single core when testing memory, so the performance is misrepresented. The latest, patched version of Sisoft Sandra XIII SP1 does support Phenom CPUs and records the results more accurately.
Real World Effects
We tested this effect on multi-threaded heavy tests to see if unganging the memory made a different in real world situations, in a mixture of tests where the Phenom achieved good and bad scores to see if there were further improvements.
-
Ganged Memory
-
Unganged Memory
Time in Seconds (lower is better)
-
Ganged Memory
-
Unganged Memory
Time in Seconds (lower is better)
-
Ganged Memory
-
Unganged Memory
0
250
500
750
1000
1250
1500
Time in Seconds (lower is better)
Despite there being a significant difference in the synthetic test results of SiSoft Sandra, there's actually very little difference in real world testing. Obviously, the benefit of having an extra memory controllers to use is offset by the fact they have less bandwidth. Either that, or one or both is being mismanaged.
Want to comment? Please log in.