orthos small fft not test all c2d cache?

[ziddey]

I'm running into a little bit o a pickle. Initially, I set vcore to 1.59v to see if system was stable at 8x450=3600. I ran blend test for 24+ hours and it was fine. Actual load vcore was somewhere around 1.52 or something if I recall correctly.

Then I dropped vcore down to 1.5v set in bios, for 1.44v under load and did small fft test. It went for 27 hours without issue.

But the system wasn't really too stable. 3dmark loops would crash it. Programs would randomly crash. I tried large fft and it'd always crash around 2 hours or so. So I bumped vcore up to 1.55v for 1.46v under load and passed 14 hours of large fft. I was about to cancel it and say it was all good, when the computer took a :banana::banana::banana::banana: with machine_check_exception. Ok.. Now I'm up at 1.575v (getting real close to original 1.59, agh). Testing large fft again, guess I'm going for full 24hr this time.

So my question is, how come small fft passes everytime? Does it not span all the l2 cache of c2d's?

I know it's said that large fft tests ram too. But my ram used to be able to run at 500 4-4-4-12 at 2.3v easily. That said, it can't do it anymore for some reason, so I'm at 450 5-5-5-15 2.3v. And when I did 24 hours of blend test with this setting (and only 2.1v) it passed fine. I also memtested for about 16 hours, and looped test5 for about 3 hours and it was absolutely fine.


edit: Also, I had another question, probably more suited for posting in the mods section but..
I have both a rev1 and rev2 ds3. On the rev1, I did the vdroop mod, which completely removed droop. However, with the rev2, droop isn't too too bad, although it's still there. However, there's another issue with the rev2. It consistently is way lower than the set vcore. For instance, I'm at 1.575v set in bios, and am getting about 1.52v idle and 1.49-50v under load. I take it that's not droop. Anyone know of any ink/pencil mods to address that? The circuitry was changed from rev1 to rev2, so the droop mod isn't the same. I'm not too good at reading the writing on the board, so I'm not sure where exactly the resistor for droop moved to (I can think of two possible spots). Maybe I'll take a picture sometime and post it

[typo_pl]

When you run the Blend Test, you have to make sure that the Prime95 numeric routines are actually doing work - if you see the red "PAGING" sign appear, then your CPU is spending time paging data in/out of memory (lotsa disk I/O, not much CPU work) and not calculating. I'm going to try and make this more evident in the next version.

Small FFTs means the FFT data *fits* in the L2 cache, not *ALL* of L2 cache is used. This is the Prime95 code for selecting the FFT size for the small FFT case (doesn't seem to have changed from v23.81 to v25.1):

if (CPU_L2_CACHE_SIZE <= 128) m_maxfft = 8;
else if (CPU_L2_CACHE_SIZE <= 256) m_maxfft = 16;
else if (CPU_L2_CACHE_SIZE <= 512) m_maxfft = 32;
else m_maxfft = 64;

If we assume that fft size of 64 would fit in a 1MB L2 cache, then a C2D with 2-4MB of L2 cache would have less than half of the L2 cache used.

Small FFTs stress the CPU (mostly the FPU) the most.
Large FFTs allocate 8MB of memory for the FFT data and stresses the CPU and that 8MB RAM mostly.
Blend/Custom >8MB allocate more than 1 set of FFT memory data and stress the CPU and a particular set of FFT memory per iteration.

So...
Small FFTs Test is like jumping up/down on the exact same spot in the kitchen of a house.
Large FFTs Test is like jumping up/down in a room & the kitchen (running to the rooms as necessary) of a house.
Blend/Custom Test is like jumping up/down in most/several rooms & the kitchen of a house.

Any damage that you incur during this jumping is your problem. :)

J


I'm running into a little bit o a pickle. Initially, I set vcore to 1.59v to see if system was stable at 8x450=3600. I ran blend test for 24+ hours and it was fine. Actual load vcore was somewhere around 1.52 or something if I recall correctly.

Then I dropped vcore down to 1.5v set in bios, for 1.44v under load and did small fft test. It went for 27 hours without issue.

But the system wasn't really too stable. 3dmark loops would crash it. Programs would randomly crash. I tried large fft and it'd always crash around 2 hours or so. So I bumped vcore up to 1.55v for 1.46v under load and passed 14 hours of large fft. I was about to cancel it and say it was all good, when the computer took a :banana::banana::banana::banana: with machine_check_exception. Ok.. Now I'm up at 1.575v (getting real close to original 1.59, agh). Testing large fft again, guess I'm going for full 24hr this time.

So my question is, how come small fft passes everytime? Does it not span all the l2 cache of c2d's?

I know it's said that large fft tests ram too. But my ram used to be able to run at 500 4-4-4-12 at 2.3v easily. That said, it can't do it anymore for some reason, so I'm at 450 5-5-5-15 2.3v. And when I did 24 hours of blend test with this setting (and only 2.1v) it passed fine. I also memtested for about 16 hours, and looped test5 for about 3 hours and it was absolutely fine.


edit: Also, I had another question, probably more suited for posting in the mods section but..
I have both a rev1 and rev2 ds3. On the rev1, I did the vdroop mod, which completely removed droop. However, with the rev2, droop isn't too too bad, although it's still there. However, there's another issue with the rev2. It consistently is way lower than the set vcore. For instance, I'm at 1.575v set in bios, and am getting about 1.52v idle and 1.49-50v under load. I take it that's not droop. Anyone know of any ink/pencil mods to address that? The circuitry was changed from rev1 to rev2, so the droop mod isn't the same. I'm not too good at reading the writing on the board, so I'm not sure where exactly the resistor for droop moved to (I can think of two possible spots). Maybe I'll take a picture sometime and post it

[ziddey]

Thanks for the clarification! So small fft's isn't enough to determine cpu stability, which explains why I was able to pass 27+ hours.

So what I'm getting is that blend stresses the most? I was able to pass 24+ hours once on blend at a higher vcore. I'll have to try that again.