Release CPU stress test

[alucart1984]

So pandoras are 50% faster than datasheet specs? (500 vs 780) @ opp3

Where does that extra performance come from?

Does overclocking to 780MHz do (long term) horrible things to pandoras?

Not sure about your first or second question but I would assume it is safe to say No to your third question.  I have had my Intel Core i5 2500K OC'd to 5.0Ghz since I got it and have had no issues.  You just have to make sure the unit doesn't get to hot.  So on a day when it's 100+ Degrees Fahrenheit outside you wouldn't want to have your Pandora overclocked and outside for long.  Your ambient temperature greatly affects stability of computers.   

 

[_wb_]

The hotter and faster a cpu gets, the more likely it becomes that noise or weird quantum effects or whatever flip a bit somewhere, which may have no consequence at all or completely crash your system. If it really gets way too hot it could burn and be damaged beyond repair.

Running your Pandora at whatever speed the cpu stress test considers safe (probably best to keep some margin, e.g. my unit can do 999MHz but I usually clock it at 900 or 950MHz) at OPP5 will not harm your unit in any meaningful way. Yes, the expected lifetime of your cpu will be a bit shorter than if you would use it at 125MHz at OPP1, but the difference in practice would be small, since 1) those expected lifetimes are quite long, so most likely some other component of your Pandora will fail first anyway, and 2) if you're using recent powersaving kernels, your cpu is essentially powered off whenever it is not used. So if you're doing something for 10 hours that takes 80% of the cpu at 500MHz and only 50% of the cpu at 800MHz, then in the first case it is as if you have taken away 8 hours from some high expected lifetime X, while in the second case you take away 5 hours from some lower expected lifetime Y.

 

[dosenfisch]

I've tested my 1GHz Pandora. At 1190MHz it's stable for 4h, with 1.2GHz it will freeze after ~30 min.

Mostly, I get "page fault" or "illegal instruction" errors followed by a freeze. The standard calculation errors with slow step down in frequency are quite rare and will be followed by a freeze after a few iterations, never reaching a stable frequency. Changing OPP between 3 and 4 makes no difference at all (not even 5MHz).

Maybe it's some kind of timing problem within the chip, starting to affect it at 1.2GHz.
If wikipedia is right, the ARM A8 architecture is designed for ~1GHz.
If they didn't optimize the design with the 45nm shrink, I'm not surprised we gain only +250MHz in comparison to the old 65nm chip.

 

[sepulep]

for rebirth, stopped after 2hr stable:


@OPP3: 847


@OPP4: 935


@OPP5: 1011

 

[Binky]

That's a good OC-er you've got there

 

[klapse]

This is a NEON-based stress test that tries to automatically find your highest stable overclock while NEON is active.

I wonder if you could release source or a build for N900? I am curious about getting numbers for my phone.

 

[urjaman]

The thing does overclocking in a propably-pandora-specific way (I dunno how N900 kernel does things TBH), but the pnd contains the source code.

EDIT: lol I ninja'd notaz by seconds (propably, Posted time is identical to minute), sorry

 

[notaz]

The sources are in .pnd itself (just unsquashfs it).

 

[White Demon]

CC Pandora reporting in.

I ran the test overnight, starting at 800mHz @ OPP5.

It climbed up to 958mHz before reporting an error, and for the last 90 minutes has been running without error at 921Mhz.

So I guess that's my limit.

I may run the test next weekend at OPP4 and see what I get, as I have run my Panda at 833mHz at that OPP level previously without any evident problems, would be interesting to see what the stress test reports.

 

[Linux-SWAT]

I tested many units, i have to say again that one whole week testing per OPP were needed in the vast majority of my tests.

 

[Binky]

 Roughly how much did their speed drop after the first few hours?

Is it safe to run the test for 4hours, look at final cpu speed x, then set x-5 as the default speed?

 

[Linux-SWAT]

It's random.

No.

 

[Binky]

Randomnesses normally operate within a range, don't they

I'm fairly certain that x-50 would be safe.

 

[Linux-SWAT]

I wrote down MHz frequently when i did my first serious tests, but gave away as my best thing to do was to simply wait ^^.

Also, i tested units per groups so i can't recall anything.

X-5 or x-50 ? ^^

 

[_wb_]

If the stress test doesn't find an error in 2 hours, even though it will find an error after 2 days, you're at a clock speed that is technically not stable, but in practice it will be very unlikely to crash at that speed. If you give your cpu a bit of rest (e.g. don't run the stress test, but "only" something like ps1 emulation, where you probably pause the game every hour or so, and even in the intense moments it'll stress the cpu less than the stress test), it'll be less likely to mess up. Also, not every cpu mess up will cause a crash - if you're lucky, the only harm done is a few pixels having the wrong color in one frame.

If you wait long enough, some error will be found at some point at any clock speed, even if only caused by some rather unlikely but not impossible bitflip caused by cosmic background radiation.

I recommend letting the stress test run overnight, check in the morning what speed it's at, and then round that down by subtracting 2% or so. E.g. on my unit, the stress test doesn't find errors at 999MHz in 8 hours or so (haven't tried for longer yet), so I set the max clock speed to 980MHz to be sure.

One factor that probably has some influence is the temperature: the higher, the worse it will perform. So it's a good idea to run the stress test while you're charging your unit - the charging will increase the temperature of your unit, so the results will be more reliable.

 

[Asiyura]

It's been one hour since the test has stopped at 1205MHz @ OPP4 on my GHz pandora.
The power is plugged in and the pandora is on the sofa (not the best thermal conductor) probably one of the worst situation in terms of heat dissipation.

I didn't get one of the best overclockable CPU, but it's already quite nice!
Anyway I never had to overclock my pandora. I think that even 750MHz would have been enough, if I wasn't using Amiga 1200 emulator or playing Homeworld

EDIT: So, the test is still at 1205MHz. I can that, without a doubt, my pandora is perfectly stable at that speed 

So it ran for about 5 hours before detecting an error… pretty much stable enough imho.

 

[aidenn]

Gravedigging, since my 1GHz Pandora is awesome: 1309* 1308** 1305*** 1300. Hell yeah.

* crashed after an hour unfortunately.

** might write 1300 next time - because I already tested it as stable - instead of editing this post over and over; and 5 MHz is nothing to write home about anyway. ;p

*** aaand I just did.

 

[Linux-SWAT]

At which OPP ?

 

[aidenn]

OPP4, 1GHzs can't do OPP5.

 

[Linux-SWAT]

Yeye, but could have been OPP3.

For instance, my OP reach the same MHz with both.