V24.12 release candidate 3

[dsouza123]

Have any of the CPU/MB/SYSTEM temps increased with 24.12 ?

Has the room the PC is in increased in temp since about the time of 24.12 ?

Are the CPU heatsink and fan coated with dust ?

If you run the previous version again with the same exponent does it
produce errors now ? Was it 32 or 64 bit ?

Are there power issues, summer time can have brownouts/voltage drops
and other power line issues also air conditioners can draw excessive power
especially when starting ?

Have any components been changed or added to the PC ?
( A new video card for example. )

[sHORTY]

System is watercooled on a battery backup to prevent against surges/brownouts.

Shrug, I'll revert back and see if I have any problems.

[Cruelty]

I have had a similar problem starting with 24.11, however it turned out to be a BIOS + RAM issue in my case (new Venice core on DFI NF4 motherboard).

[db597]

Originally Posted by sHORTY

Odd that the problem doesn't show up until 24.12, however.

Ran for for over 2 months straight with previous builds and no problems. I will run some tests and see what I can find, however.

Do you mean previous builds of 24.12 or previous builds like 24.11 and lower? Each version of Prime95 gets more optimised and faster. So I suspect that the newer builds and putting more and more stress on the hardware - especially the memory. The 3% "speedup" of going from 24.11 to 24.12 could be what's pushed it over the limit.

[sHORTY]

24.11 worked 100%, that's what I am referring to.

[g0ods]

Any idea how long before V24.12 is released onto the GIMPS download/free software page?

[Prime95]

Originally Posted by g0ods

Any idea how long before V24.12 is released onto the GIMPS download/free software page?

The non-optimal Celeron D timings has thrown a monkey wrench into things. I'd like to get one more Celeron D special benchmark, but none have been forthcoming.

Assuming I need to have different FFT implementations for the Celeron D and Willamette then I'll come out with version 24.13 RC1 next week.

[g0ods]

Originally Posted by Prime95

Assuming I need to have different FFT implementations for the Celeron D and Willamette then I'll come out with version 24.13 RC1 next week.

Clealry the new version is faster on my P4 over 23.8, but out of interest I checked on my 600MHz Duron and most FFTs were about 9% quicker. Would that be the use of the 3DNow code?

Miles

[kjaget]

Originally Posted by Prime95

Good find. I looked at the web site and didn't find this.

I'm trying to figure out why your Celeron D and my Willamette are getting different optimal FFTs. Your 4-way and my 8-way is one difference (implying prime95 is not spreading the FFT data in the L2 cache optimally). Yours is probably DDR memory and mine is Rambus, so I have longer latencies. Yours is a Prescott core and mine is Willamette so it could be due to architectural changes.

If the problem is due to FFT data layout in the L2 cache then it could have implications for other CPUs too - making this the most serious problem I've looked into with the 24.12 release candidates.

And here you were worried we weren't finding enough problems

WCPUID shows the following for L2 -

256KB size, 4-way, 64 byte line, dual sectored

The rest of the cache info, in case it makes a difference :
Trace cache - 12K uops, 8-way
ITLB shows 4KB pages, 4-way, 32 entries
next line (?) 4KB pages, 2M/4M byte pages, 128 entries (not sure what this refers to)
L1 - D side 16K, 8-way, 64 byte line, dual sectored
DTLB - 4KB pages, 4Mbyte pages, 64 entries

The system has 1GB of DDR running at 195MHz.

There's a utility that comes with CPU-z (http://www.cpuid.org/cpuz.php) which calculates cache latency. The results show 4 cycles for L1 (D-side I assume) and 29 for L2. This seems to match what others have measured for Prescott (e.g. http://www.xbitlabs.com/articles/cpu...escott_8.html), and is way higher than the 2 & 19 cycles for Northwood. I'm guessing the changes in the L2 are a big part of the performance difference beween the two cores, but who knows what else has also changed.

One thing I plan to do is rerun the test at stock speed (2.53 instead of 3.7). Hopefully the relative speeds stay more or less the same, but I'm starting to worry a bit that there's something subtle going on with the overclock. I've been getting good results for a while, but there are never any guarantees.

[kjaget]

Originally Posted by kjaget

One thing I plan to do is rerun the test at stock speed (2.53 instead of 3.7). Hopefully the relative speeds stay more or less the same.

The non-OC results are consistent with the OCd results. There were a few cases where the best parameters for the OC vs non-OCd chips were different, but most of them are cases with either very small or very large FFT sizes. For the small FFT sizes, two fastest results were within a few hundredths of a msec - a tie. This was the case in 192K (normal and all-complex) and 384K.

8192K and 12288K are the two large FFTs where the fastest time was different for OCd and non-OCd chips. The speed difference between 1st and 2nd place in OCd results was about .8% (8192K) and .07% (12288K), so again - a tie.

For the rest of the results, the same params give the best results for both OCd and non-OCd chips.

These results also fit in between the 2.4 and 2.6 GHz results from the other thread, so you should have some good info to go on.

Kevin

[CADavis]

Will my Celeron 2.4GHz run better with 24.12 that with 23.8?
Also I have a Celeron 733MHz, which version should I run?

Neither of these are Celeron D.

Also, my Athlon 64 3500+ runs about 25% faster now thanks!