Linpack Challenge

[lopri]

Inspired by n7's overclocking experiment here (whooping 50 loops! 61 GFlops!), I decided to run Linpack one more time on my 955 BE system. While most here know Linpack is an excellent stability test, less known is it also is quite an important benchmark in HPC world. Linpack scores will show up in most server CPU reviews, though not in desktop CPU reviews.

I don't know how Linpack scores translate to performance in desktop applications, but we're enthusiasts, and we don't take kindly to real-world stuff anway - so why not check out how great our desktop CPUs are in this benchmark? Also don't forget that Linpack is one of the most stressful tests on CPU-memory subsystem, so if you've overclocked your CPU it's worth spending some time to make sure things are stable.

So let the fun begin. Let me loosely outline the rule so that we can gather some results that are meaningful. First, grab LinX. It's a GUI-enhanced version of Linpack and it will automatically run either 64-bit version or 32-version upon detecting your OS. (It also tells you if your system fails to finish the runs) Then,

1) Use all available memory. A system from a cold-boot will let Linpack have 80%+ of total system RAM.
2) Dual-core CPUs - at least 10 loops, Quad-core CPUs - at least 20 loops.
3) Report the Max GFlops achieved. LinX will give that number at the bottom.
4) Report the OS, CPU frequency, memory frequency as well as anything else you think that matters. ^_^

Edit: Added a Google Doc link so that anyone can enter the data. (Don't mess with others' plz)

Linpack Challenge Entries

Here is mine.

CPU: Phenom II 955 BE
Core Frequency: 3600 MHz (1.35V, air-cooled by Zalman CNPS9700)
NB/Uncore Frequency: 2250 MHz (1.10V)
Memory Frequency: DDR2-900 (5-5-5-15)
Peak Performance: 45.33 GFlops

 

[cmdrdredd]

System in sig. Peak GFlops 49.0164

 

[Markfw]

it stopped after 22 seconds with error. But I run F@H 100% load for 3 weeks now, and no reboot, EUE, or any problems...

 

[Griswold]

Case in point why F@H isnt a good stability indicator then. You probably still give them alot of garbage WUs back with that rig...

This Linpack really makes my box steam like no other tool. I'll probably do a full run myself later this weekend.

 

[Idontcare]

Case in point why F@H isnt a good stability indicator then. You probably still give them alot of garbage WUs back with that rig...

This Linpack really makes my box steam like no other tool. I'll probably do a full run myself later this weekend.

silent data corruption

 

[Markfw]

Case in point why F@H isnt a good stability indicator then. You probably still give them alot of garbage WUs back with that rig...

This Linpack really makes my box steam like no other tool. I'll probably do a full run myself later this weekend.

Their software is designed to find this. Exactly why you get EUE's. If you are not a F@H person, you wouldn't understand. The point is, I DON"T, and OCCT and prime95 all say its fine. This has been documented with other people.

 

[Rubycon]

Their software is designed to find this. Exactly why you get EUE's. If you are not a F@H person, you wouldn't understand. The point is, I DON"T, and OCCT and prime95 all say its fine. This has been documented with other people.

Does it err with default clock speed? If not there is a change and output from that machine should not be trusted. Now if YOUR application does data checking and is OK then this COULD indeed be OK. It simply means your specific application is not stressing and/or using a feature of the CPU instruction set that is explicitly FAILING at higher clock speeds.

 

[masteryoda34]

54.0742 peak. (Only 3 passes, I don't feel like running 20.)

 

[BTRY B 529th FA BN]

o i thought it said the 6 pack challenge, my bad

 

[cmdrdredd]

I should rerun, I notice I didn't use the larger size option to run when I did my test so 20runs took 6min. If I get around to it I might leave it run again using all memory. I don't know, playing Dragon Age on that machine which is very addicting.

 

[zagood]

Went through 5 passes and got bored. You can only watch a gray bar flash for so long.

The e5200 is really not meant for apps like this. Peak performance 17.7430 GFlops @ 3.5ghz. I'll run a full 10 passes tonight.

One thing I love about these stress tests is how it clears out your ram - before test, 1.7gb in use, after test, ~800mb!

p.s. it reads my CPU speed @ 4.163 GHz (actual is 3.5ghz - 10.5x333). There are some other programs that do this as well in Win7, wonder what's up?

 

[Hyperlite]

38.1 Gflops
Athlon II X4 @ 3.2ghz
2GB OCZ Gold @ 1660ish 8-8-8-30 1.7v

i was at 32gflops at stock clocks (2.6ghz) with 4gb of ram, but i have to RMA it due to a faulty stick.

29gflops @ stock clocks/2gb ram...so i should be able to break 40 Gflops when i get my RAM back next week.

 

[lopri]

I should rerun, I notice I didn't use the larger size option to run when I did my test so 20runs took 6min. If I get around to it I might leave it run again using all memory. I don't know, playing Dragon Age on that machine which is very addicting.

Also try running with HT off. I remember reading in an AT review that Linpack is such a well optimized application for multi-core systems that HT actually hurts the performance.

 

[Markfw]

Does it err with default clock speed? If not there is a change and output from that machine should not be trusted. Now if YOUR application does data checking and is OK then this COULD indeed be OK. It simply means your specific application is not stressing and/or using a feature of the CPU instruction set that is explicitly FAILING at higher clock speeds.

Well, I know F@H is picky... So I shut it down this time. It ran find in 5 passes. Doesn't like to run at the same time as F@H. 47.5475 max gflops...

That was with HT on. I turned it off. Now I get 52.0341 gflops ! It doesn't like HT.

@ 3.9 (195 x 20) ghz it does 53.4008 gflops..

Lets try for more !

 

[lopri]

Added a Google Doc link in the first post so that anyone can enter the data. (Don't mess with others' plz)

http://spreadsheets3.google.com/ccc?key=tXI1w9SelPpsT6BbGXFvcrw&hl=en

 

[Idontcare]

Well, I know F@H is picky... So I shut it down this time. It ran find in 5 passes. Doesn't like to run at the same time as F@H. 47.5475 max gflops...

That was with HT on. I turned it off. Now I get 52.0341 gflops ! It doesn't like HT.

@ 3.9 (195 x 20) ghz it does 53.4008 gflops..

Lets try for more !

mark what do your L2$ miss and success rates look like when you compare HT on versus off while running linpack?

(run perfmon while running linpack w/HT and w/o HT and see if the HT case is cache starved, I am just curious)

 

[Markfw]

I am installing win7 on that box now. I see most people here are running a 64 bit OS and version .0.6.3, while I was in 32 bit XP, and 0.6.4. I need to get on a more even playing field. Posting from another box....

 

[JFAMD]

Mine wins:

http://www.wired.com/gadgetlab/2009/11/most-powerful-supercomputer/

 

[Idontcare]

Mine wins:

http://www.wired.com/gadgetlab/2009/11/most-powerful-supercomputer/

1.759 petaflops

Ouch...we're all playing around with our sticks and knives and you show up with the "nuclear option" in your arsenal!

 

[nyker96]

Mine wins:

http://www.wired.com/gadgetlab/2009/11/most-powerful-supercomputer/

I can't believe they upgraded 37k+ CPUs that along would take months. good thing AMD designed the 6cores as a drop in replacement for the socket, no need to upgrade cooling solutions.

but it says: "AMD says it will offer the first eight-core and 12-core x86 processors for high performance computers early next year." I have never heard of this been said anywhere. If it can be done 8/12 build on 45nm would require lot more cooling.

 

[luv2increase]

I would like to let everyone know that the GFlops reported by LinX is inaccurate. There were many people at the EVGA forums in the i7 4ghz club who had different values across the board which didn't make sense.

Also, I get almost 10GFlops more when tested in SiSandra than compared to LinX. Yes, I'm talking about version 0.6.3 as well.

 

[Markfw]

Win7 64 bit, new number...55.6358 !!!

 

[Idontcare]

I would like to let everyone know that the GFlops reported by LinX is inaccurate. There were many people at the EVGA forums in the i7 4ghz club who had different values across the board which didn't make sense.

Also, I get almost 10GFlops more when tested in SiSandra than compared to LinX. Yes, I'm talking about version 0.6.3 as well.

I would trust linx over sisandra when it comes to a robust implementation of linpack.

If folks are getting variable numbers from their OC'ed rigs then it sounds like their systems are getting throttled (either temp or over-current) or they are incurring excessive ECC requests during processing and it is dragging down their scores.

 

[n7]

Added my results lopri was referencing in the first post to the Google spreadsheets thing.

Also adding my numbers from my Q9650 system.

CPU: i5 750
Core Frequency: 4158 MHz (1.375v - Noctua NH-U12P)
NB/Uncore Frequency: 3168 MHz
Memory Frequency: DDR3-1980 9-9-8-24-1N 1.6v
Peak Performance: 61.03 GFlops





CPU: Q9650
Core Frequency: 4104 MHz (1.256v - TRUE w/ 2x Noctuas)
Memory Frequency: DDR2-1140 5-5-5-15 1.8v
Peak Performance: 58.75 GFlops



Man i miss my Q9650...well, it's still here, just sitting in a box waiting for me to sell it.
Really should do that, but it was such a good gal, hate to see her go...

 

[lopri]

CPU: Athlon X2 7750 BE
Core Frequency: 3200 MHz (1.325v, Stock Opteron HSF)
NB/Uncore Frequency: 2400 MHz (1.325V)
Memory Frequency: DDR2-1066 (5-5-5-15-2T)
Peak Performance: 20.24 GFlops



Edit: Didn't know I was running an older version of LinX. I don't think it'd make a difference but I guess I'll double check the channel log of the latest version.