E8400 OCing Curiosity

[Kermit de frog]

Perhaps you should try what BaldEagl's saying and reduce your memory speed to achieve a 1:1 ratio between your memory and FSB.  In my opinion, unless you increase to 2:1 ratio, you won't see any improvement except in synthetic benchmark programs.  Either your memory is failing that your OC'd settings or your "North Bridge" is failing at the OC'd memory setting.  Perhaps it's due also to your higher mobo temps of 44C.  In my opinion, running below 100F would be ideal. 

I run an E6400 2.13GHz OC'd to 3.03GHz
PC6400 underclocked from 400MHz to 380MHz (4 CL) (Dual Channel configuration to saturate FSB)
FSB is at 380MHz (effectively 1520MHz and up from 1066MHz)

[katanaso]

Interesting concept, Kermit, and I read that having one faster than the other (FSB or Memory) can create a bottleneck.

It would be easy enough to drop the memory down to 1:1, allowing me to reduce the Northbridge voltages.

I wonder about a 533 FSB with 7.5 multiplier, then a 1:1 ratio as well.

The system is stable now, but yeah, it's warm.

[Skuzzy]

You guys are forgetting about the bus masters in your computer, such as the video card and hard drives.  The FSB has a direct impact on the transfer performance to/from RAM.

It does translate into real world performance gains.  In my class in the videographers test, my computer is the fastest, even though there are thjose that have their CPU's clocked higher.  The videographers test is a real world rendering test, not a synthetic benchmark.

[BaldEagl]

Interesting concept, Kermit, and I read that having one faster than the other (FSB or Memory) can create a bottleneck.

It would be easy enough to drop the memory down to 1:1, allowing me to reduce the Northbridge voltages.

I wonder about a 533 FSB with 7.5 multiplier, then a 1:1 ratio as well.

The system is stable now, but yeah, it's warm.

Generally, to achive the same outcome, if you lower the multi with a higher FSB you'll create more heat than running a higher multi with a lower FSB.  I can't see any reason to ever drop the multiplyer.  If you're going to do that just drop the FSB instead and the vcore with it and eliminate heat.

[katanaso]

Generally, to achive the same outcome, if you lower the multi with a higher FSB you'll create more heat than running a higher multi with a lower FSB.  I can't see any reason to ever drop the multiplyer.  If you're going to do that just drop the FSB instead and the vcore with it and eliminate heat.

Thanks.  I'm reading about it as I can to see the pros and cons, and if it's even worth it.

I'm having to run my NB a little higher than standard as is, but that's only since I went to 4 sticks of RAM.  When I had 2 sticks, voltages were default for the NB.

Is the 1:1 ratio for CPU/RAM a non-issue on motherboards that allow various ratios like the one I'm using, where I can get 1066 (533x2) and above by picking a specific ratio as needed?

[BaldEagl]

Is the 1:1 ratio for CPU/RAM a non-issue on motherboards that allow various ratios like the one I'm using, where I can get 1066 (533x2) and above by picking a specific ratio as needed?

No, the issue is in syncing the clock speeds so that the RAM and CPU are at the same frequecies and both are exchanging data on the same clock cycles.  Besides that overclocking the RAM beyond it's manufacturer specs and what's actually needed can destabilize the RAM resulting in an unstable OC.

If you could sync the RAM to the CPU at a lower clock speed it would be more stable.  Then you could try taking the timings from 5-5-5-15 to 4-4-4-12 or something similar and actually speed the RAM up by lowering latencies.  That would still leave the CPU and RAM on the same clock cycles but speed the data handling internally for the RAM so that it might be able to respond to a CPU data request in one instead of two clock cycles for instance.

[katanaso]

If you could sync the RAM to the CPU at a lower clock speed it would be more stable.  Then you could try taking the timings from 5-5-5-15 to 4-4-4-12 or something similar and actually speed the RAM up by lowering latencies.  That would still leave the CPU and RAM on the same clock cycles but speed the data handling internally for the RAM so that it might be able to respond to a CPU data request in one instead of two clock cycles for instance.

Interesting.

If I could do that, with my current settings, it would be running the RAM at 900MHz.  (450x2).

So, it would be 900MHz 4-4-4-12 or the current 1080MHz 5-5-5-15.

I wonder what the difference would be, giving up speed for a decrease in latency.

I'm not sure the memory could do it though.

[633DH98]

I'm assuming not all MBs will let you change memory latencies?  I can't seem to find it in the bios for my Gigabyte GA-M74GM-S2 rev 2.0.  Only options I can find are Memory Clock and DDR2 Voltage.

[BaldEagl]

Interesting.

If I could do that, with my current settings, it would be running the RAM at 900MHz.  (450x2).

So, it would be 900MHz 4-4-4-12 or the current 1080MHz 5-5-5-15.

I wonder what the difference would be, giving up speed for a decrease in latency.

I'm not sure the memory could do it though.

You're not actually giving up speed.  As it stands your RAM is running a faster clock cycle than the CPU and has to wait for the CPU on each data exchange effectively slowing it to the CPU's clock speed.

Just as with OCing the CPU reduce latencies one at a time and one step as a time and test for stability after each change.  Don't just jump to 4-4-4-12.

I'm assuming not all MBs will let you change memory latencies?  I can't seem to find it in the bios for my Gigabyte GA-M74GM-S2 rev 2.0.  Only options I can find are Memory Clock and DDR2 Voltage.

Not sure.  I can change mine.  I've got an eVGA 780i motherboard.  I don't remember offhand where it's located in the BIOS though.

I know that my Kingston HyperX DDR2 800 RAM defaults to 5-5-5-15 timings anytime you add a new stick even though the spec is 4-4-4-12.  Kingston does this to assure compatability with a wide range of motherboards so, to get to spec, you have to manually set the timings and voltages.  That being the case I'd guess that theres a place to re-set timings somewhere in your BIOS.

[katanaso]

You're not actually giving up speed.  As it stands your RAM is running a faster clock cycle than the CPU and has to wait for the CPU on each data exchange effectively slowing it to the CPU's clock speed.

Just as with OCing the CPU reduce latencies one at a time and one step as a time and test for stability after each change.  Don't just jump to 4-4-4-12.

Gotcha.

What are the drawbacks from the RAM running faster than the CPU? (540 vs 450).  More latency, due to the timings of 5-5-5-15, but anything else?

I lost my internet access yesterday, so I didn't see your post until today, else I would've played with the memory settings last night.

Thanks again.

[Skuzzy]

May I suggest a review of the block diagram in this Intel PDF file.

It is pertinent to all Intel CPU's prior to the i7 family, regardless of the Intel motherboard chipset.  I'll point out the memory bus speed is between the memory controller and the RAM.  The clock rate between the memory controller and the CPU is the external clock rate of the CPU.

Only reason I am bringing this up is due to some misinformation in this thread.  Regardless of the FSB or CPU clock rates, the CPU always has to wait on a memory read on any cache miss.  The CPU does not have to wait on writes as those are buffered in the memory controller which actually does the RAM writes for the CPU.

[katanaso]

Thanks, Skuzzy.  That makes it plain to see.  I appreciate the diagram!

Is there an explanation as to why it's nearly 50% faster from MCH to RAM than MCH to CPU?

[Skuzzy]

That will be dependent on the memory controller in the chipset and the CPU being used.  The clock between the memory controller and the CPU is the external CPU clock.

The FSB clock is between the memory controller and the system RAM.

In early chipsets there was only one clock source and various multipliers and dividers were used to generate the FSB and CPU/MCH clocks.

With the Core Duo release the chipsets implemented completely independent clock sources for the RAM and CPU.

The ratio you see between the FSB and CPU clocks is for the memory controller so it can properly sync the two buses.

[katanaso]

Gotcha.  Thanks for the explanation there.

For resuming some tweaking, I was researching various boards, and read a bunch about tightening specific memory settings.  I went over the the Mushkin Forums and found some good info there, and I changed some settings that lowered the latency by over 10ns, as well as increasing the read and write throughput several percent to around 9500 MB/Sec or so.

The FSB is 450 (9 multiplier for 4050MHz) and RAM is still 5-5-5-15 at 1080MHz (540x2).  I had some attempts at 4-4-4-12 at 800MHz, but it wouldn't POST.  Searching various sites suggested it's harder to work with 4 sticks of RAM than with just 2, but I wasn't going to open the system last night.  Interestingly enough, I could overclock the RAM up to 5-5-5-15 at 1200MHz and boot into Windows, run programs, and work.  However, it wasn't as fast as when I was able to modify the specific settings of tRRD, tWTR, tWR, tRFC, tRTP, and Static tRead at 1080MHz.

The system has been stable, and it passed some stress tests (Memtest86 and Intel Burn Test), and it is noticeably faster.

I stayed up too late fiddling with things, so I can't post the tweaks now, but I will tonight, as well as make sure it's running stable.

This is definitely an interesting and fun learning experience.

[Chalenge]

With the e8400 and Corsairs 800 Mhz RAM I linked and synched like the gurus suggest on the eVGA forums which required setting the memory voltage to factory specs and lowering the timing (from 5-5-5-18-2 to 4-4-4-12-1). Once thats all done it looks like the memory will under-perform at 667 but in fact with the RAM operating at 800 (1600 FSB manual entry) the CPU is operating at a stable 3.6 Ghz (3.599). The way it was explained to me is that Intel wants the market so even if they run out of 8400 processors they will take the under-achieving 8600s and label them as 8400s. Im not sure if thats true but every 8400 I have seen can easily do 3.6 and on air they dont even approach getting 'hot.'

Everest Home Edition reports my memory latency is zero (0) which cannot be right.