Intel Burn Test Slower with XMP

[jlwmanagement]

I recently got a new mobo: MSI Z77A-GD65 and i easily got my 2500-k to 4.5Gz. The problem is that I initially left XMP off for stability testing and was getting 104 GFlops. I then enabled XMP and was getting 96-98 GFlops. I never got a crash and am stable with Prime either way. I was just wondering why would I see such a result?

Any ideas?

Thanks in advance!

 

[Idontcare]

I have noticed this too, and it seems to go away if you bump up the cpu voltage a smidge.

While I have no proof, my interpretation here is that the cpu is turning in lower GFlops because it is on the hairy edge of an instability but the ECC circuits on the L1$ are catching the errors and the cpu is recalcuting the computation in question (and eventually gets it right) such that the error doesn't manifest into an actual error at the application level.

But the recompute time comes at the expense of fewer net calcs done per second, so lower GFlops rating.

 

[jlwmanagement]

Thanks for the input. I have my CPU on 1.390 Volts simply because I did not have time to dial it down before my vacation was over so I am not sure that it is getting enough juice. Although in your defense, all other voltage options are on auto except for vDroop which is 100%

 

[FDCPCZ]

Gflops is always lower with HT enabled.

IDC
You mis interpreted what the OP was saying.
Your explanation of lower reported Gflops is correct in the context of edge of stability scenarios, howeverit is not relevant in the OP's case.

 

[Puppies04]

Gflops is always lower with HT enabled.

IDC
You mis interpreted what the OP was saying.
Your explanation of lower reported Gflops is correct in the context of edge of stability scenarios, howeverit is not relevant in the OP's case.

I love the fact that you are trying to tell someone whos little finger has forgotten more about pcs than you will ever know that they misinterpreted something while simultaniously claiming that XMP has something to do with HT and also apparently claiming that a 2500K even has HT.

Brilliant.

 

[jlwmanagement]

Indeed the 2500-K has 4 cores and no HT

 

[john3850]

Will HT lower your scores
Now my new 3370k @4400 untweaked is only hitting 96 GFlops on 8 threads.
While tweaked 2500k@4600 with the same memory is getting 110+GFlops with 4 threads.

 

[Puppies04]

Will HT lower your scores
Now my new 3370k @4400 untweaked is only hitting 96 GFlops on 8 threads.
While tweaked 2500k@4600 with the same memory is getting 110+GFlops with 4 threads.

Yes, it has been noted that HT off can produce more gflps than HT on in IBT.

 

[john3850]

Yes, it has been noted that HT off can produce more gflps than HT on in IBT.

Now my new 3370k @4400 untweaked is only hitting 96 GFlops on 8 threads.
While tweaked 2500k@4600 with the same memory is getting 110+GFlops with 4 threads
HT off 3370k @4700 is now 115.1867

 

[Puppies04]

Now my new 3370k @4400 untweaked is only hitting 96 GFlops on 8 threads.
While tweaked 2500k@4600 with the same memory is getting 110+GFlops with 4 threads
HT off 3370k @4700 is now 115.1867

As I said it has already been noticed on several occasions in this forum. You would need someone like IDC to tell you where the stolen gflps end up though, my guess would be intel is stealing them and selling them on the black market.

 

[Ferzerp]

Now my new 3370k @4400 untweaked is only hitting 96 GFlops on 8 threads.
While tweaked 2500k@4600 with the same memory is getting 110+GFlops with 4 threads
HT off 3370k @4700 is now 115.1867

Your 3770k @ 4.4 Ghz will get much better scores if you use the proper thread count. HT only helps when your cores are waiting. With Linpack, it never is, so you have a penalty from HT instead of a benefit. It helps in lmost all cases, but.... Linpack is one of the few exceptions. You don't have to turn it off, just use the proper thread count.

Just lower it to a thread count equal to your cores.

 

[Makaveli]

I love the fact that you are trying to tell someone whos little finger has forgotten more about pcs than you will ever know that they misinterpreted something while simultaniously claiming that XMP has something to do with HT and also apparently claiming that a 2500K even has HT.

Brilliant.

I don't know about you but I laughed!

 

[Puppies04]

I don't know about you but I laughed!

yes I had a good chuckle at that one, had to read it a couple of times (and re-read the OPs post) to make sure it was as dumb as it sounded.

 

[Idontcare]

Gflops is always lower with HT enabled.

IDC
You mis interpreted what the OP was saying.
Your explanation of lower reported Gflops is correct in the context of edge of stability scenarios, howeverit is not relevant in the OP's case.

We may be talking about different things altogether?

The OP didn't change their processor's HT settings (it doesn't have HT), all he reportedly changed was the ram settings in the BIOS.

The faster ram may be feeding the cpu cores at a less rate-limiting pace now, and as such the cores are now working just a smidgen harder to process data (restrictor plates are removed ) but because they are working a tad harder now they are also operating in the very narrow band of "stable only by virtue of L1$ ECC saving your butt".

But even that is just a wag (wild ass guess) on my part. But I have noticed this behavior on my 2600k.

If I drop the Vcc too low, but not so low as to cause crashing, then my GFlops value declines but IBT doesn't throw any errors. Raise the Vcc just one notch or two and the GFlops climbs back to where I'd expect it to be on a per-GHz level.

 

[john3850]

Your 3770k @ 4.4 Ghz will get much better scores if you use the proper thread count. HT only helps when your cores are waiting. With Linpack, it never is, so you have a penalty from HT instead of a benefit. It helps in lmost all cases, but.... Linpack is one of the few exceptions. You don't have to turn it off, just use the proper thread count.

Just lower it to a thread count equal to your cores.

With HT on 3370k @4700 I set it to 4 cores ran and got 108-114 GFlops.
HT off 3370k @4700 is now 114-115 GFlops.

Close enough to leave HT on thanks Ferzerp.

 

[GammaLaser]

We may be talking about different things altogether?

The OP didn't change their processor's HT settings (it doesn't have HT), all he reportedly changed was the ram settings in the BIOS.

The faster ram may be feeding the cpu cores at a less rate-limiting pace now, and as such the cores are now working just a smidgen harder to process data (restrictor plates are removed ) but because they are working a tad harder now they are also operating in the very narrow band of "stable only by virtue of L1$ ECC saving your butt".

But even that is just a wag (wild ass guess) on my part. But I have noticed this behavior on my 2600k.

If I drop the Vcc too low, but not so low as to cause crashing, then my GFlops value declines but IBT doesn't throw any errors. Raise the Vcc just one notch or two and the GFlops climbs back to where I'd expect it to be on a per-GHz level.

I don't think this is quite the reason, mission-critical CPUs such as Itanium/POWER are the only processors which have the soft-error detection in the core datapath itself and can retry instructions when an error is detected. (source, page 13)

If it's Vcc correlated, maybe a more plausible explanation is the error happens on the L1$ fill side, and when the CPU tried to fetch this data it is incurring L1 miss due to error in the L1.

 

[Ferzerp]

With HT on 3370k @4700 I set it to 4 cores ran and got 108-114 GFlops.
HT off 3370k @4700 is now 114-115 GFlops.

Close enough to leave HT on thanks Ferzerp.

np. What you're seeing with HT on with 4 threads is the penalty for the threads hopping between cores. If you turn HT off entirely, there will only be 4 so they won't really do that.

 

[FDCPCZ]

That will teach me not to read the thread properly.

 

[tweakboy]

I have noticed this too, and it seems to go away if you bump up the cpu voltage a smidge.

While I have no proof, my interpretation here is that the cpu is turning in lower GFlops because it is on the hairy edge of an instability but the ECC circuits on the L1$ are catching the errors and the cpu is recalcuting the computation in question (and eventually gets it right) such that the error doesn't manifest into an actual error at the application level.

But the recompute time comes at the expense of fewer net calcs done per second, so lower GFlops rating.

IdontCare nice one, classic IdontCare style...... your post powns, and ya ECC is only for servers, not consumer and ECC ram is expensive so is a xeon mobo and cpu ........ I think todays desktop is stable enough to crunch those numbers without error.......that is all it is

 

[sm625]

Assuming XMP means "extreme memory profile"... I remember back in the days of multiple memory ratios some site did some tests that showed how performance unexpectedly decreased at certain ratios even though the actual speed was faster.

 

[jlwmanagement]

We may be talking about different things altogether?

The OP didn't change their processor's HT settings (it doesn't have HT), all he reportedly changed was the ram settings in the BIOS.

The faster ram may be feeding the cpu cores at a less rate-limiting pace now, and as such the cores are now working just a smidgen harder to process data (restrictor plates are removed ) but because they are working a tad harder now they are also operating in the very narrow band of "stable only by virtue of L1$ ECC saving your butt".

But even that is just a wag (wild ass guess) on my part. But I have noticed this behavior on my 2600k.

If I drop the Vcc too low, but not so low as to cause crashing, then my GFlops value declines but IBT doesn't throw any errors. Raise the Vcc just one notch or two and the GFlops climbs back to where I'd expect it to be on a per-GHz level.

Thanks IDC! I did some tweaking last night and got mine up to 4.7 on 1.385V. I will try for 4.8 and I hope to get it stable around 1.35V tomorrow. I always thought VCC = Vcore. Did you mean VCCIO? (which I thought was for RAM voltages) Thanks again!

 

[jlwmanagement]

Yep, upping the VCCIO did it. Thanks IDC!

 

[Idontcare]

Yep, upping the VCCIO did it. Thanks IDC!

:thumbsup: You can thank me but you're the one that did all the work, good job and glad to hear you nailed it!