Gigabyte GA-MA78GM-S2H overclocking memory of its own will?

[jack_aubrey]

I recently started having stability issues with my ca. month-old GA-MA78GM-S2H system. Running memtest86 turned up a few hundred errors on one module; reseating it in a different socket helped for a bit, then I had to pull it. Yesterday the one remaining module started acting up as well. Finally I paid closer attention to the memtest86 screen and noticed that my DDR2 800 memory was running at more like DDR2 846! Keep in mind this is using all the automatic/default settings on the motherboard, with an Athlon X2 5000 Black Edition CPU. After fiddling with manual settings I've dropped the CPU speed a bit and got the memory clock down to 801, which seems to be addressing the problem (cross fingers).

This was with el cheapo Crucial Ballistix Tracer 1GB modules (those flashing LEDs are really annoying, but it was cheap, and qualified for the mobo according to Gigabyte's website), so it's not overly surprising that the overclock led to instability. But why on Earth would the motherboard decide of its own accord to overclock my memory beyond its own specs? It hardly seems "optimized" for the BIOS to run memory beyond what either the memory or the motherboard specs claim it can do.

[dukla2000]

But why on Earth would the motherboard decide of its own accord to overclock my memory beyond its own specs?

It is not the mobo - it is the X2 5000 (remember the AMD memory controller is in the CPU). The memory bus speed is set by a divider from the CPU speed. In this case your CPU is running 2600MHz and has (based on parameters from the BIOS about your memory and no doubt from the memory itself) decided to use a divider of 6, ergo 433MHz. Being DDR this nets to 866MHz. If you set the BIOS to 667MHz the CPU would accordingly decide to use a higher multiplier and also avoid the problem.

Legend has it that the divider logic would always underclock your memory if it could not get the exact speed so it is curious that in your case it is overclocking. I would have expected (in your case) a BIOS setting of 800MHz to cause the CPU to use a divider of 7 = 371MHz = DDR 742. It may be that Gigabyte are trying to be clever (for benchmarking purposes?) and thus telling the CPU that the memory is faster than it really is. Do you get the same symptoms if you force a DDR 800 setting in BIOS instead of (presumably) AUTO?

[jack_aubrey]

Legend has it that the divider logic would always underclock your memory if it could not get the exact speed so it is curious that in your case it is overclocking. I would have expected (in your case) a BIOS setting of 800MHz to cause the CPU to use a divider of 7 = 371MHz = DDR 742. It may be that Gigabyte are trying to be clever (for benchmarking purposes?) and thus telling the CPU that the memory is faster than it really is. Do you get the same symptoms if you force a DDR 800 setting in BIOS instead of (presumably) AUTO?

Thanks for the insight. With default settings

CPU clock ratio: Auto
CPU Freq: 200 MHz
Set Mem Clk: Auto

memtest86 reports:

Athlon 64 X2 2611 MHz / RAM: 435 MHz (DDR 870)

When I change to

Set Mem Clk: Manual
Mem Clk: DDR800

memtest86 shows the same results, so manually forcing DDR800 in and of itself doesn't help. However, if I also set

CPU clock ratio: x12

then I get

Athlon 64 X2 2411 MHz / RAM: 801 MHz (DDR 803)

Which is how I'm running now. If I set

CPU clock ratio: x13

then it's back to DDR870 and ignoring the manual mem clk setting.

It's also a bit puzzling that in the first case memtest86 applies the DDR2 multiplier in what it reports (435 MHz / DDR 870) and in the second does not (801 MHz / DDR 803). But that's a side issue.

[dukla2000]

I'm just wondering how accurately Memtest is reporting things, although based on your initial errors it seems likely your memory was overclocked.

Reason being I just did a first fire on my new system: Abit A-S78H, 4850E and cheapo DDR 800 RAM. With Optimised BIOS settings MemTest 86+ said I was running my memory at 416 = DDR 832. (The 4850 is 2500MHz) So it seems that any mobo 'cheating' is not confined to Gigabyte. (Then again I imagine the mobo should just be reporting facts for the CPU to decide, like anything the CPU doesn't get direct from SPD.)

IMHO - I prefer playing with clocking by varying the HTT. That way you should still be able to use CnQ (because CnQ wants to play with the multiplier but desists if it is non-standard to start with).