Rundown to SuperPrime 1.0Final : Need your help!

[KiD0M4N]

Hey people,

This is the latest beta of the SuperPrime benchmark.

Home page: www.superprime.org

Direct download: Links and Downloads Manager - SuperPrime™ (you can download ZIP file directly.)

Requirements:

- Microsoft .NET 2.0 Runtime (x86 and x64) (or if you are feeling generous, install .NET 3.5 - recommended)
- Microsoft VC++ 9.0 Runtime (you can download this from within the program as well)
- A fast system!

A example screenshot:



Please give me your feedback so that I can improve this tool and make it as usable as possible.

Shoutout to all the quad owners: Care to give this a spin? Show us where you stand.

Regards,KiD0M4N

[KTE]

You've got a first tester.

I've tested it to spot problems. What's happening at 2200 vs 2400 multi thread?

Test 1&2 @ 2200MHz - 400MHz - 4-4-4-8 1T: single thread and multi-thread





Test 3&4 @ 2400MHz - 436MHz - 4-4-4-8 1T: single thread and multi-thread

[KiD0M4N]

Damn photobucket.... I cant see any of the images because it says that "Bandwidth exceeded." Please rehost the images at www.imgx.org or www.imageshack.us or something

I will be glad to give out fixes for any/all problems you see.... and thanks again for testing... (waiting for the rehosted pics now :P)

You've got a first tester.

I've tested it to spot problems. What's happening at 2200 vs 2400 multi thread?

Test 1&2 @ 2200MHz - 400MHz - 4-4-4-8 1T: single thread and multi-thread





Test 3&4 @ 2400MHz - 436MHz - 4-4-4-8 1T: single thread and multi-thread

[mAJORD]

I can see the images on this comp, but not on my other one.. same photobucket error. weird.

to cure your curiosity Kidoman the problem is:

all 1E8:

2.2ghz 1thread: 3.35
2.2ghz 4thread: 1.12m

2.4ghz 1thead: 3.15m
2.4ghz 4thread: 1.5m

[KiD0M4N]

I can see the images on this comp, but not on my other one.. same photobucket error. weird.

to cure your curiosity Kidoman the problem is:

all 1E8:

2.2ghz 1thread: 3.35
2.2ghz 4thread: 1.12m

2.4ghz 1thead: 3.15m
2.4ghz 4thread: 1.5m

Thats very weird.... maybe a one of thing?

I have run this benchmark on a 8 core machine (na, not Clovertown, but a Xeon server I had access to at work) and it scaled perfectly.

Could you please try a rerun of the same, maybe it was something in the background.

[[XC]Atomicpineapple]

Heres my results.

[KiD0M4N]

Hey mAJORD,

Were the CPU specs detected properly in the screenshots? TIA

[KiD0M4N]

Heres my results.

Thanks for posting the result.

One small doubt: Was anything running in the background during the benchmark? Also, could you please post a screenshot of taskmanager showing the utilization of all the cores during the entire duration of the 8 core run.

Remember, there is a single threaded part to the benchmark (I put it in because I wanted to include a single threaded memory intensive test which put in some weight to single threaded performance of the CPU.)

TIA, waiting for the screen shot

PS: First time SuperPrime has run on a 8 core clovertown machine

[KTE]

I can see the images on this comp, but not on my other one.. same photobucket error. weird.

to cure your curiosity Kidoman the problem is:

all 1E8:

2.2ghz 1thread: 3.35
2.2ghz 4thread: 1.12m

2.4ghz 1thead: 3.15m
2.4ghz 4thread: 1.5m

Thanks. Yep all that is correct. Sorry full bandwidth ran out in 3 days.

Here's the links:
http://img223.imageshack.us/img223/6744/superpak7.png
http://img223.imageshack.us/img223/3...rpmultibz7.png

http://img223.imageshack.us/img223/1...erp2400am0.png
http://img223.imageshack.us/img223/8...00multimf6.png

BTW your program (IE8) is very good for RAM stability testing. Much better than SuperPi 32M.

[KiD0M4N]

Thanks. Yep all that is correct. Sorry full bandwidth ran out in 3 days.

Here's the links:
http://img223.imageshack.us/img223/6744/superpak7.png
http://img223.imageshack.us/img223/3...rpmultibz7.png

http://img223.imageshack.us/img223/1...erp2400am0.png
http://img223.imageshack.us/img223/8...00multimf6.png

BTW your program (IE8) is very good for RAM stability testing. Much better than SuperPi 32M.

Thanks for your words.

I am gonna get cracking on AMD datasheet for its Barcy processors so that I can detect the cache sizes properly. The intel algorithm isn't working out too well evidently.

Also one request to you KTE (since you are the first guy to run Phenom on this :P)

Could you post a screen shot with the task manager showing the utilization of all the 4 cores and run a 4 core IE8? it will help me diagnose the anomaly in the results.

Remember, the last part of the benchmark is intentionally single threaded, to give weight to single thread performance (majority of apps in real world) as well as the multi threaded performance.

[KTE]

Could you post a screen shot with the task manager showing the utilization of all the 4 cores and run a 4 core IE8? it will help me diagnose the anomaly in the results.

Here you go



Some pointers:

It uses 88-97% CPU usage throughout. Stays from 92-94% most of the time.

Single thread section loads ALL cores 25% each. I can change this by setting affinity obviously.

Also, there's a problem with it. If you cancel it early then it takes that time as your benchmark time.

[Metric]

Q6600 @ 3.6GHz (400x9.0). Side note: I'm interested (as it seems many prime benchmarks are popping up), what algorithm are you using to calculate primes?

[[XC]Atomicpineapple]

Heres the screenies you asked for, all background tasks shut off.
Final times were 24.XX sec and 1.39 min.

[KiD0M4N]

Here you go

Some pointers:

It uses 88-97% CPU usage throughout. Stays from 92-94% most of the time. The threading efficiency goes down slightly with the number of cores I guess. With my C2D E6600, the CPU utilization stays pegged at 100% during most of the run.

Single thread section loads ALL cores 25% each. I can change this by setting affinity obviously.

Also, there's a problem with it. If you cancel it early then it takes that time as your benchmark time.

The threading efficiency goes down slightly with the number of cores I guess. With my C2D E6600, the CPU utilization stays pegged at 100% during most of the run.

When in the single thread section, the OS does the division of work... so to optimize your score, set the affinity :P

If you cancel prematurely, the score appears with a brown under line and will reflect the same when saved to the hard drive (did you try that feature btw :P)

Q6600 @ 3.6GHz (400x9.0). Side note: I'm interested (as it seems many prime benchmarks are popping up), what algorithm are you using to calculate primes?

Nice score.... I am using "Coarse multi threading" instead of fine grained multithreading using a combination of primality algorithms ATM. The particular algorithm which is used is decided based on the "hugeness" of the number in question. And then some trickery to keep all the cores loaded to the hilt at all times

Heres the screenies you asked for, all background tasks shut off.
Final times were 24.XX sec and 1.39 min.

Thanks for the screen shot mate

So it is indeed loading all the cores I wish there was someone with > 8 cores in XS so I could test it at greater than 8 threads

A new version, 0.9.7 is coming out momentarily with updates to read cache information from AMD cpus.

[KiD0M4N]

Q6600 @ 3.6GHz (400x9.0). Side note: I'm interested (as it seems many prime benchmarks are popping up), what algorithm are you using to calculate primes?

Actually, I developed SuperPrime in Oct'06. After Nov'06 kinda lost touch because of work and then wPrime popped up.... but anyways, messa back, and committed to making this a helluva benchmark :P

[KiD0M4N]

New version (0.9.7) out people.

Includes AMD cache detection and couple of improvements. Please report back if it works as expected.

http://forums.erodov.com/local_links.php?catid=7

[KTE]

I've tried it before with P4, Celeron D and E6750. It loads Intel cores to 97-100% while loading AMD to 88-95%.

When in the single thread section, the OS does the division of work... so to optimize your score, set the affinity :P

4 cores @ 1.8GHz, NB @ 1.96GHz, HT @ 1.47GHz, RAM @ 700MHz

4 threads - 4 cores


4 threads - 2 cores


2 threads - 2 cores


2 threads - 1 core


1 thread - 1 core


1 thread - 2 cores

If you cancel prematurely, the score appears with a brown under line and will reflect the same when saved to the hard drive (did you try that feature btw :P)

Can't tell, I'm color blind

A new version, 0.9.7 is coming out momentarily with updates to read cache information from AMD cpus.

The dll you're using reads K10h speed wrongly. It'll do that for a while TBH because the speed is calculated in complex ways. Why not use the same method wPime uses to get CPU info?

[KiD0M4N]

Damn KTE! Thats some extensive testing :P :P Appreciate it

Now, the DLL which I am using is written by me from scratch since my search for a CPUID library failed miserably. It seems the real time CPU clock on the X4 is operating at the NB's frequency and not the core frequency. Interesting. (thats why the NB frequency is showing instead of the Core frequency, just a educated guess.)

Also slightly confused with your "x threads - y cores" results. For example, in the "2 threads - 1 core" screenshot, have you disabled 3 cores from the X4 and run the benchmark with the Core - 2 option? The 4 thread - 4 core result really intrigues me. Please post that screenshot (as mentioned below to help me debug this.)

I have updated my code to read the cache of AMD cpus now, care to give 0.9.7 a spin and please (if possible) take a screenshot of a 1E8 run (4 cores) + taskmanager (showing the 4 cores/utilization) + CPUz screenshot. You can use the "Screenshot" button on SuperPrime itself to easily save the screenshots

0.9.7: http://forums.erodov.com/local_links.php?catid=7

[KTE]

NB speed is what dictates CPU speed although it doesn't calculate it, if NB speed is too high it'll drop the CPU and HT speed right down to compensate. It's quite complex.

Also slightly confused with your "x threads - y cores" results. For example, in the "2 threads - 1 core" screenshot, have you disabled 3 cores from the X4 and run the benchmark with the Core - 2 option? The 4 thread - 4 core result really intrigues me. Please post that screenshot (as mentioned below to help me debug this.)

Let me clarify.

Cores = how many cores I enabled for the benchmark to use by setting Affinity (works fine).
Threads = how many cores I chose in SuperPrime.

SuperPrime ran 4 threads on 2 cores faster than 4 threads on 4 cores.

BUT 2 threads on 1 core was slower than 2 threads on 2 cores (as expected)

Meaning, it has problems with over 2 cores.

CPU utilization stays as I said earlier for ALL runs.: 88-95%.

If I set affinity to 1 core with 1 thread, 2 threads, 3 threads or more, it'll stay at 88-95% on average.

If I set affinity to 4 cores with 1 thread, 2 threads, 3 threads or more, it'll stay at 88-95% on average.

I'm at work and have brought the system in but dep. head looks a little mad right now so can't access the system to start the tests with new build yet.

[KiD0M4N]

This behavious is quite weird... I will read up on the docs more so that I can code a different speed detection algorithm for the 10h family.

As stated here (http://www.xtremesystems.org/forums/...&postcount=13), the algorithm is scaling perfectly with the increased cores (upto 8 in that post, dual clovertown) but it seems weird with the Barcelona.

Waiting for your screenshot of the 4 thread 4 core run so that I get a idea on whats wrong and a validation of my cache detection code

PS: If you can point me to "any" tech doc which will aide in CPU Mhz calculation, I will be grateful.

NB speed is what dictates CPU speed although it doesn't calculate it, if NB speed is too high it'll drop the CPU and HT speed right down to compensate. It's quite complex.

Let me clarify.

Cores = how many cores I enabled for the benchmark to use by setting Affinity (works fine).
Threads = how many cores I chose in SuperPrime.

SuperPrime ran 4 threads on 2 cores faster than 4 threads on 4 cores.

BUT 2 threads on 1 core was slower than 2 threads on 2 cores (as expected)

Meaning, it has problems with over 2 cores.

CPU utilization stays as I said earlier for ALL runs.: 88-95%.

If I set affinity to 1 core with 1 thread, 2 threads, 3 threads or more, it'll stay at 88-95% on average.

If I set affinity to 4 cores with 1 thread, 2 threads, 3 threads or more, it'll stay at 88-95% on average.

I'm at work and have brought the system in but dep. head looks a little mad right now so can't access the system to start the tests with new build yet.

[KTE]

Windows is very accurate at getting Phenom frequency real-time and so is the RDTSC. If you want to grab that, use the Windows API Counters QueryPerformanceFrequency() and QueryPerformanceCounter().

I can't read this because I have no PDF software installed, but try it (attached).

I ran your latest BETA

[KiD0M4N]

I was looking throught wPrime and it seems its using a embedded CPUz program (from www.cpuid.com) to do the detection. I can do that as well, but thats not a very elegant solution. I will keep looking.

PS: Looked at your attachment, I went through that exact document (order# 25481) to code the cache detection algorithm. I think I am going to try the CPUz approach tonight :P

And I will definitely wanna work with to see why its utilizing only 88% cpu. Weird :/ On Intel quads, it uses 100% very nicely.

[Metric]

Getting a benchmark error with the 5E5 set (Both 0.9.6 and 0.9.7 are affected) Rest of the benchmark sets seem to be working.

[KiD0M4N]

Getting a benchmark error with the 5E5 set (Both 0.9.6 and 0.9.7 are affected) Rest of the benchmark sets seem to be working.

Fixed that bug in the latest version

http://forums.erodov.com/local_links...tid=7#linkid15

http://www.superprime.org/

Changelog:

Version 0.9.8

- Added CPUz integration.
- Fixed a small bug in L1 cache detection for AMD processors.
- Improved (or not?) the threading algorithm to improve CPU utilization.
- Changed version to 0.9.8

[Movieman]

Sorry for the slow bench..All I have is a server machine with those horrible FBDimms that everyone talks about..