Multi-Threaded Super-Pi Contest

poke349

New Member
LOL, no biggie. I'm still third place in the 256m and second place in the 512m. I'm sure some I7's and some high clocked C2Q's will put me down LOL. I'm hoping to improve my scores some once I install my new motherboard and TRUE cpu cooler to try to hit 4.0ghz stable. Even 3.9ghz gets toasty with the GeminII cooler, about 68c.

Any decent dual-socket board will also do... :rolleyes:
But it wouldn't be fair if I joined in. (since I know all the tweaks)

*I'll be glad to share them though. :D


And if anyone can beat the times here:

http://www.numberworld.org/y-cruncher/#FastestTimes

Lemme know, I'll be glad to update it as a new world record.

The times I use there are the "computation times", not the "total time including writing digits".
I do it this way because it throws out the effect of disk speed. (since it's a CPU/ram benchmark)
 

G25r8cer

Active Member
Srry but could you change my stats again?

Im running a "Phenom II 955be"

Not be picky or anything but, I dont want to give other people a different impression/thoughts
 
Any decent dual-socket board will also do... :rolleyes:
But it wouldn't be fair if I joined in. (since I know all the tweaks)

*I'll be glad to share them though. :D


And if anyone can beat the times here:

http://www.numberworld.org/y-cruncher/#FastestTimes

Lemme know, I'll be glad to update it as a new world record.

The times I use there are the "computation times", not the "total time including writing digits".
I do it this way because it throws out the effect of disk speed. (since it's a CPU/ram benchmark)

Here's some new runs, earlier I was only running my ram at 1333 but now these runs are with the ram at 1790.

2m- .598
http://i554.photobucket.com/albums/jj422/smercer1000/new2m598.jpg

4m- 1.193
http://i554.photobucket.com/albums/jj422/smercer1000/4m1193.jpg

8m- 2.352
http://i554.photobucket.com/albums/jj422/smercer1000/8m2352.jpg

16m- 4.611
http://i554.photobucket.com/albums/jj422/smercer1000/16m4611.jpg

32m- 9.472
http://i554.photobucket.com/albums/jj422/smercer1000/new32m9472.jpg

64m- 20.425
http://i554.photobucket.com/albums/jj422/smercer1000/64m20425.jpg

128m- 45.596
http://i554.photobucket.com/albums/jj422/smercer1000/128m-1.jpg

Btw Poke you wouldn't have any tips to give would you?:p
 
Last edited:

poke349

New Member



Nice results!!! I'll update them later today. ;)

Hmm tweaks...

Hardware/Software Tweaking:
  1. For anything less than 15 million or so... disable HT.
  2. On Core 2, get your FSB as high as possible, period. This program needs a LOT of bandwidth. (2x Xeon X5482 @ 3.2 GHz + 64GB DDR2 @ 800 MHz 1600 FSB is faster than 2x Xeon X5470 @ 3.33 GHz + 128GB DDR2 @ 667 Mhz 1333 FSB for computations over 50m.)*
  3. Once your memory speed reaches a certain point (Core i7), it doesn't help much to go any higher. In that case, try lower frequency and tighter timings.
  4. The program is quite sensitive to uncore frequency.
  5. Windows Server HPC gives you control over a ton of thread and scheduling-related stuff. These can give a huge speedup on small computations where the programs spends a lot of time creating and destroying threads. (I think some of the other versions Windows may have it too.)
  6. Try multiple runs. The smaller the computation, the more inconsistent the timings are (due to threading).
  7. For this thread, write the digits to a ram disk. Since the times used for this thread include that final write to the disk which is disk-limiting.
  8. For Core i7, keep your temps below 80C. That's about the point where turbo gets throttled. For my machine, for large computations (> 250m), 4.3 GHz runs slower than 4.2 GHz because 4.3 is just enough to push temps past 80C - which disables turbo and lowers the clock to 4.1 GHz.
  9. For Core i7, use Windows 7 (or Server 2008). It knows how to deal with HT better than all previous versions of Windows.

*Sorry, I don't have any less extreme examples.
The only Core 2-based system I have is the dual X5482s with 64GB of ram. (in my sig)
The closest matching system I found benchmarks for is dual X5470s with 128GB of ram. (credit to Shigeru Kondo from Japan)

Program Tweaking:
  • Minimize the window.
  • Use the batch mode. It lets you loop benchmarks. It's complete with validation so results are just as valid as benchmark mode.
  • Batch mode gives you full control over the # of threads to run. This lets you override what the programs chooses. Sometimes it might be worth running more threads than cores. And for the really small runs, decreasing threads may help.
  • Try a different binary. The program chooses what it "thinks" is the fastest for your machine. But this isn't always true. For example: The "x64 SSE4.1 ~ Nagisa" binary runs faster than the "x64 SSE4.1 ~ Ushio" binary for anything less than ~16 million digits on Core i7.
  • Should memory be a tight squeeze, run the program a few times and kill it a few seconds after it sustains 100% cpu. This will force the OS to page out enough stuff to avoid thrashing. (This is crucial to getting 2.5b to run efficiently with 12GB of ram.)
 
Last edited:
Thanks man for all the information, I really wasn't expecting that. I am going to play around with the uncore frequency as well as my ram some more. I am trying to find a sweet spot with the timing and speed but my ram is limiting me. I really need to get some Corsair 2000 mhz stuff. I noticed that the top 1m was using a different binary and the best so far with the ushio for me is like .295 so I will also try re running all my sub 16m with the other binary.
 

G25r8cer

Active Member
Got one for the 256m standings

High temp of only 42c

Capture-57.png
 

Jet

VIP Member
Hey all!

I updated the scores to not include the time to write to disk to mesh better with the official WR records that do not include that. The only change in placement is between bigrich0086 and 87dtna--87dtna got dropped to 5th in the 32M.
 

poke349

New Member
Hey all!

I updated the scores to not include the time to write to disk to mesh better with the official WR records that do not include that. The only change in placement is between bigrich0086 and 87dtna--87dtna got dropped to 5th in the 32M.

Updated the WR list. :D Sorry it took a while... was almost up the whole night in a Lan Party... :eek:

Also, the frequency that the program gives is almost always wrong when the multiplier isn't stock. (So it also won't detect turbo boost increases.)
So CPUz screenshots will be helpful.

It's very hard to detect the right frequency... hence why CPUz exists. But I keep it there to aid with validations.



Alex
 

G25r8cer

Active Member
87dtna: Could you post a full cpu-z screenshot? I seem to have hit a brick wall with fsb and could use some help.
 

87dtna

Active Member
87dtna: Could you post a full cpu-z screenshot? I seem to have hit a brick wall with fsb and could use some help.

I only used the multiplier to overclock. Black edition! Multiplier is simply at 19.

What ghz are you hitting a wall at? And whats your CPU voltage at? Whats your cpu NB and HT set at?
 

G25r8cer

Active Member
I only used the multiplier to overclock. Black edition! Multiplier is simply at 19.

What ghz are you hitting a wall at? And whats your CPU voltage at? Whats your cpu NB and HT set at?

I get a bsod with 19x multi and everything else on auto. Vcore is running between 1.408-1.424

Bsod is when running ycruncher
 

Jet

VIP Member
I get a bsod with 19x multi and everything else on auto. Vcore is running between 1.408-1.424

Bsod is when running ycruncher

When overclocking, Auto is your enemy. It changes stuff when you don't want it to be changed, and you can never track what adjustments actually work. Get everything off manual, and then work from there--and keep a log! :)
 

G25r8cer

Active Member
When overclocking, Auto is your enemy. It changes stuff when you don't want it to be changed, and you can never track what adjustments actually work. Get everything off manual, and then work from there--and keep a log! :)

Raised the HT and Vcore and coretemp wont give me temps

Edit: Got her stable at 3.8ghz now

Just gave it somemore voltage and set the ht back to normal

What next? Bump the multi or drop ram and raise fsb?
 
Last edited:

Jet

VIP Member
Raised the HT and Vcore and coretemp wont give me temps

Edit: Got her stable at 3.8ghz now

Just gave it somemore voltage and set the ht back to normal

What next? Bump the multi or drop ram and raise fsb?

As long as your ram is at/under stock, you're fine.

Do you have an unlocked multi? It's really either way--keep raising the FSB until you run into motherboard issues. You'll know it when you need to start raising your NB voltage.
 

Jet

VIP Member
Yes I have unlocked multi

Just wanted to know which way is better

The higher the FSB, the higher the overall system performance, but it will raise temperatures more in general due to the Ram and NB being overclocked in addition to the CPU. Personally, I'd raise the FSB until you hit a wall, and then play around with the multi.
 
Top