Device Bandwidths, Memory Configuration and Bottlenecks?

[Veritas3]

I have a Gigabyte motherboard (X38-DQ6) which officially supports a FSB of up to 400 MHz (although it can easily and reliably go as high as 475 MHz). So setting the FSB to 400 MHz means that my Intel Core 2 Duo E8500 which normally runs at 333MHz is overclocked and with a multiplier of 9.5 means it will "sing" at 3800 MHz rather than at a stock 3.16 GHz. All fine so far.

I also have 4 GB of Crucial DDR2-1066 (PC2-8500) RAM which can run at either 800 MHz (5-5-5-18) at 1.8v or 800 MHz (4-4-4-12) at 2.0v or 1066 MHz (5-5-5-15) at 2.0v.

From what I have seen with benchmarks (UserBench Encode 2009), they are directly related to increases in CPU speed with barely significant increases noted after tightening memory timings or increasing memory frequency.

To date, I haven't looked for a correlation between benchmark results and pure increase in FSB speed (whilst keeping CPU speed fixed by way of the multiplier), so I'm unsure what raising the FSB speed to 450MHz would do in that regard?

So my question is, what is the bottleneck in such a system (FSB, CPU, RAM?) and what configuration(s) should I implement for the memory with regards to RAM:FSB divider, frequency, timings and voltage for

1) best performance and 2) most stable and least heat generation (two different objectives)?

Because in theory, memory run at 533 MHz can process 8.5 GB/s, compared to memory run at 400 MHz which can only transfer 6.4GB/s so running the RAM at 533 MHz should be a good thing right?

Thing is, it makes a barely measurable performance gain in the benchmarks.

Any advice very welcome and if you know of any good articles on this topic, I would really appreciate any pointers.

Thanks

 

[Edrick]

When I has a socket 775 system, I always ran my memory at a 1:1 ratio. I had a E8200 then a Q9300, both overclocked to 400Mhz FSB. So my memory always ran at 400Mhz with very tight timings. I found that to be the best for my setup.

But as I soon found out, the memory controller on the oder socket 775 chipsets was probably the biggest bottleneck. Moving to a P55 chipset made a noticable difference in memory speeds (double according to Everest).

Old system = (dual channel P45) DDR2 @ 400mhz 5-5-5-15 = 6.2GB/s (1:1 ratio)
New system = (dual channel P55) DDR3 @ 600mhz 7-7-7-19 = 12.3GB/s (1:4 ratio)

 

[lopri]

http://www.anandtech.com/show/2427/4

http://www.anandtech.com/show/2458/4

 

[betasub]

Bottleneck under what workload? Benchmarks like the "UserBench Encode 2009" you mention?

 

[Veritas3]

Thank-you for your replies. They have been most informative. I read those articles on FSB and memory thanks lopri, although a little heaving going in places.

I have decided to lower the OC to 3600MHz for the E8500 which allows me to set the VCore in the BIOS to "Normal" rather than the 1.26875 that I had to set it at when OC'd to 3800MHz.

I have experimented with various memory timings and bench-marked in SiSoft Sandra, which shows that memory running at 533MHz (5-5-5-15) gives better bandwidth than memory running at 400MHz (4-4-4-12) - no surprises there, but in real world testing which "UserBench Encode 2009" seeks to emulate, there's very little difference (53.4 v 53.0).

My worry at the moment though, is that when the Vcore is set to normal in the BIOS, it shows under CPU-Z as 1.09v under no load (C1E & EIST function kicks in and lowers the multiplier to 6x400), 1.216v under light load at 9x400; and 1.2v under load (vdroop).

Is 1.2v too low for an E8500 OC'd to 3600MHz? I have set all the other voltages to normal other than the Vdimm (2.0v).

How do you test for under-voltage? Most benchmarks test for over-voltage don't they??

Thanks again for your help.

PS. Those articles lopri that you pointed me to, assume very fast memory which you can run at low CAS. For example under the "Best Pick" recommended DDR2 configurations it states a FSB of 400MHz, a tRD of 5 and a tCL (CAS) of 4 at DDR-1000.

I don't think mine would do it, and also, I can't identify which setting in the BIOS for my GA-X38-DQ6, is actually tRD?

My BIOS has "Act to Act Delayed" (tRRD currently set at 4) and "Read to Write Delay" (tRD_WR currently set at 8). Is it either of these two? I fiddled with the latter, trying 7 and 6 but this caused instability when opening too many windows whilst I was OC'd to 3800. Hence the drop back to the more family friendly 3600

 

[Edrick]

I had a Q9300 which I OC'd to 3.0Ghz (on a Gigabyte board). I set the VCore to normal and ran my tests.

It would put me at about 1.21 under full load which would bluescreen after about an hour. Under normal work however, I would never know there was an issue. I manually set my VCore to 1.23 and no more bluescreens.

So yes, the "normal" setting can under volt your CPU when overclocking on Gigabyte boards.

 

[Veritas3]

So yes, the "normal" setting can under volt your CPU when overclocking on Gigabyte boards.

Thanks Edrick for the info. So when you talk about setting Vcore to get a certain value, do you mean the figure in the BIOS or the resulting figure as reported by CPU-Z before or after Vdroop?

And I see that in the "PC Health Status" page of the BIOS there is another Vcore voltage again...

I have set Vcore in the BIOS to 1.25v (which is "normal" for a E8500 according to the BIOS), "PC Health" reports 1.268v, CPU-Z reports 1.248v no load, and 1.232v under load.

Considering that using the "normal" setting for Vcore was resulting in 1.2v under load, and I was worried about a BSOD (although nothing had happened after 6 hrs of prime95) I am hoping it should be stable now. What do you reckon?

 

[Edrick]

When I refer to my VCore values, I am referring to the cpu-z values. Naturally the setting in the BIOS would be higher. I just do not remember how many steps above I had to set the BIOS.

I also used LLC on the gigsbyte boards (old and current). I know a lot of people do not like to use it, but I think it works well. I am able to keep my VCore (in cpu-z) the same value for both load and idle.

1.232v, in your case with your speeds, seems to be very good. If you are stable at that value, I would be happy. Thats where I was with my Q9300 OC'd to 3Ghz (1600Mhz FSB). Your idle value of 1.248v seems a bit high. Maybe try LLC to bring that value down to 1.232v as well.