
Inquiring Minds ...
would like to know if it runs fine sitting in a drum of kitchen oil?
overclocked super computers here we come!
The Raspberry Pi Foundation has equipped its credit card-sized computer with an overclocking - and overvolting - mode that doesn't tear up your warranty when you activate it. The organisation is pitching the update as the introduction of a "turbo mode" which dynamically ups the core voltage and frequency but only so long as …
see how high you can go until you find a setting that renders the board unstable.
Overclockers have always puzzled me. The manufacturer knows exactly which are the critical pathways in the CPU. They can test and appropriately speed-grade their chips by exercising these pathways. Intel turbo-mode is supported, meaning that Intel has tested your CPU at the highest turbo speed they support. You'll get correct results, as long as you stay within the thermal envelope.
But if your CPU has not been manufacturer-tested at the speed you are clocking it., all bets are off. Which matters most - getting the right results, or getting wrong results faster?
It's not an entirely rhetorical question. If you are rendering frames for a movie or game, wrong results are either immediately obvious or of no significance. However, that's a special case. One bit wrong in the allocation bitmap of a filesystem, or in the compression or encryption of a datastream, and the eventual loss may be huge. Even in a render farm, overclocking has risks: it's the same overclocked CPU crunching the pixels, and adjusting the allocation bitmap of the disk on which the results are stored.
So the question: why overclock a Pi, when you can so easily buy a faster system?
OK, I'll bite. Overclocking works because the manufacturer of the chip has tested it for the highest guaranteed clock. They err massively on the side of caution in the majority of the times, and so there is always room for overhead. The tolerances involved in the manufacturing process means there is always a safe overhead (they won't build a chip that *just* operates inside the clock frequency, as a minor glitch could make it go wrong) so it's a case of figuring out where that overhead reaches you. It's trivial to find too, run soak tests until it goes wrong. Pretty much every CPU will find 10% margin and that's without tweaking voltage levels.
Where are you finding a faster system than a Pi in that price range and that form factor then?
Just like the people who "chip" the engine management units on their cars:
"I chipped my car and I got 40 more horsepower out of the engine, and a load more torques"
to which I tend to reply something along the lines of:
"But the engine management unit is there to manage the engine and prevent the engine from damaging itself and make sure it runs for a reasonable lifetime, what you've done is potentially dramatically reduce the lifespan of your engine. Those racecars which have the same engine and different software aren't exactly designed for longevity..."
They are also designed to prevent ham fisted morons from damaging the car through poor driving and understanding (like co workers who didn't realize their BMW's were rear wheel drive and why that mattered). Its horses for courses if you know what your doing you can run things to finer tolerances, if you bounce off rev limiters and mash the throttle like a gorrila then yeah taking out the protective systems is not a good idea.
i ran overclocked cpu's when i needed more power from ever aging gaming rigs in the pre 1ghz days, these days though i would just buy a better part but for something like the Pi this is a great move to allow flexibility dependent on project (i.e more speed/power consumption from the same low price unit if power is not an issue)
>But the engine management unit is there to manage the engine and prevent the engine from damaging itself and make sure it runs for a reasonable lifetime
A remap doesn't remove anti-knock and other protection for your engine. In the UK its often worth remapping from generic programmes - as the fuel is better here and air temp ranges less extreme - hence you get not just increased power but increased MPG - certainly the case with pretty much all VAG vehicles.
Places like MTM do little else and since they guarantee their work and are still in business, clearly aren't killing many cars.
Sometimes the software in the ECU is created for an entire market (EMEA for example), therefore it can be very conservative with its configuration to comply with the regulations for all countries in that area.
Also in the case of German cars there has been a horsepower threshold in the past (150 or 180 HP?) above which the insurance leaps massively. So some German made cars have been artificially limited to this bandwidth.
There's also the fact that companies don't always want their turbo diesels to be as fast as their hot hatches in 0-60, so the diesels have their performance clipped.
Many Turbo diesels are identical in construction, and a manufacturer offers a variety of power outputs at varying price points, it is usually only the ECU mapping that varies the power, so chipping for more power/torque is a relatively safe bet for lesser vehicles, but best uprate those braks as well.
"Overclockers have always puzzled me. The manufacturer knows exactly which are the critical pathways in the CPU. They can test and appropriately speed-grade their chips by exercising these pathways. Intel turbo-mode is supported, meaning that Intel has tested your CPU at the highest turbo speed they support. You'll get correct results, as long as you stay within the thermal envelope."
This isn't necessarily true: the manufacturer may choose to sell "underclocked" parts, as it's cheaper to produce everything on a single process rather than having multiple production lines. Also, their stress-tests are (presumably) based on relatively low thermal tolerances and assume that the customer is using standard voltages and a cheap OEM heatsink, rather than one of the mega-fancy liquid-cooled, silver-plated, multi-fan, mega-finned uber-heatsinks that your average overclocker likes to bolt atop their CPU.
So there's often a fair amount of milage to be had by going beyond the manufacturers recommended specs...
And in answer to "why overclock a Pi": why not?
Longevity: Statistical analysis (what all chip companies do to determine longevity) indicates there will be some decrease in lifespan of the SoC on the Raspi due to overclocking+overvolting, but not enough to cause any problems. (i.e. still 20 years plus at 100% duty cycle).
So there is no downside to using this patch.
"But if your CPU has not been manufacturer-tested at the speed you are clocking it., all bets are off. Which matters most - getting the right results, or getting wrong results faster?"
I say this in greatest respect: You're talking bollocks.
I am running a Core 2 Duo 2.66ghz system at 3.2ghz. I had to upgrade the cooling to cope but no wrong results at all. The popular system stress test program Prime95 is very handy for checking to see wether any faults happen after overclocking.
There's nothing wrong with overclocking, heck even my graphics card has a couple of overclocking settings built into the standard software.
Not true in all cases, but there are a few reasons that overclocking can be reasonably safe:
1) The manufacturer clocks are typically the highest that they can guarantee for a large batch of CPUs, meaning that a significant number can go higher without any issues.
2) Overclockers typically have better CPU cooling than the default (which the manufacturer clocks are set for), meaning that errors are less likely at higher clock speeds.
3) The same silicon is occasionally used for both low and high speed CPUs, just with a multiplier/FSB change. It is cheaper to manufacture than having separate designs. In this situation, it is usually safe to overclock the lower CPU to match or exceed the better CPU.
4) Most overclockers test the speeds they attain. If the CPU can manage 100% utilisation for e.g. 24 hours, then it is unlikely to have errors in the future.
But if your CPU has not been manufacturer-tested at the speed you are clocking it., all bets are off. Which matters most - getting the right results, or getting wrong results faster?
I'm writing this on an eeePC which runs a 1.6GHz Atom thinks-it-is-a-dual-core-but-isn't-really. If I switch to "Super" mode, the built in manufacturer-supplied tool will clock the processor at 1.8GHz; this being quite useful for rendering high bitrate HD video on this older bit of hardware. Of course, it makes more heat so it helps to hijack the fan to run it harder than the BIOS would. You might find the enemy of your filesystem isn't so much the clock speed as the excess heat.
I once saw a computer (W95 days) that (from an IDE drive dump) looked to have spat its paging file all across the start of the harddisc (wiping out the boot sector, the primary FAT, and a load of other necessary things). What went wrong was the processor fan failed. It was an early Pentium. It made a lot of heat. Big heatsink was hot enough to fry an egg (absolute fact: when the system was decomissioned I powered it up, let it get hot, dropped a raw egg on the thing). Anyway, the processor actually ran that hot (um, pre-eggy, that is) and booted a DOS game from floppy. What had failed was the SIMM stuck directly above it. I popped in a small SIMM and shielded it from the CPU heat with some alufoil for the DOS game test. The bad SIMM, in another system, failed in unpredictable ways due to memory tests.
Thus to the Pis and Beagles and similar. I think overclocking the processor won't necessarily be too traumatic to it so long as you keep heat in check. When you've maxed it, you probably won't get as far as completing a boot. I worry about the clocking of my Beagle because of the CPU and memory being sandwiched together. Heat tends to make things bend. I'd rather run slow for a long time than faster for a mere twenty years. It might sound an eternity, but the Beeb beside me is older...
"Buy it from another supplier then"
Yes. I've had one on order from RS since July, with "up to 11 weeks" given as the delivery time. I ordered one from CPC on Friday, just after 5pm, and it arrived yesterday.
"Granted it's now sitting here on a desk with nothing to do yet..."
On your desk? Mine's still in the box it was delivered in - which is much too big to put on the desk.
The Pi is built around a 700Mhz embedded SoC. It further constrained since all IO is via SD or USB.
It was obvious before launch that this thing was never going to be fast. The best way to get performance out of it is to avoid using the CPU as much as possible, keep running processes to a bare minimum, and push stuff onto the GPU or utilising the hardware enabled codecs. i.e. use it like an embedded SoC.
And the point of your post was....what?
Slow machine becomes 50% faster? In fact 700Mhz is a lot of oomph for those of us who grew up with 2Mhz BBC Micros. Even the 90Mhz Pentiums, or 400Mhz Alpha could do a lot of work if the sw was written effectively. This device is considerably faster than those machines.
"And the point of your post was....what?"
My point was to point out reality, that some people expect more of the Pi than is reasonable given what it is built around. The SoC is designed for embedded applications like set top boxes and when it steps outside that role (e.g. when people try to run a desktop on it) it's limitations become immediately clear. No need to get defensive about it, it's just a fact.
It'll probably be faster than many DEC, HP, IBM, SGI, Sun (etc) era UNIX workstations that used to run the CDE desktop.
This can readily be confirmed as CDE is freeware as of not long ago:
http://sourceforge.net/projects/cdesktopenv/
It may or may not be sensible to use it for that.
So what? It is what it is - an engineering marvel that (amongst other things) shows just how much folk are being ripped off by the manufacturers of smart TVs, PVRs, etc.
"So what? It is what it is - an engineering marvel that (amongst other things) shows just how much folk are being ripped off by the manufacturers of smart TVs, PVRs, etc."
Throw in a PSU, flash card, case, IR dongle / remote, cable, wifi dongle, and cost wise it's peeking into the same zone as media players which contain similar SoCs. Yes it's a nice device which has gathered a good community, but let's not fool ourselves into thinking it's the only thing out there.
Indeed eBay is filled with devices which contain a 1.5Ghz CPU, 4GB flash, 1GB ram, USB, wifi and HDMI and preloaded with Google TV. You'd get one for $50 if you were lucky or cut out the eBay and bought direct from the countless vendors on Alibaba. These sorts of cheap devices are becoming commonplace really.
BTW I'd never advocate ever buying a smart TV. I consider the additional expense a waste of money given that this "smart" functionality would most likely be bitrotten a few years down the line. The main advantage of a smart TV is potentially the content can overlay the picture and it could do other clever stuff such as picture scaling but not enough to convince me its worth it.
I ordered my Pi's back In March, one RS and one from Farnell.
Then somehow I messed up and ordered a second from Farnell (this may be beer related).
I received my first Pi from Farnell. My second from Farnell turned up two days after they removed their "one per customer" restriction two weeks later.
My Pi from RS turned up 6 weeks after the second from Farnell.
I gave the extra one to a family friend's son who hasn't come out of his room since!
Over 6 months is a long old time to sort out a supply problem. I think RS should just give up, they're not making any friends with that level of service.
There's no link to the source so I have to ask - was this just a simple kernel patch? Or is there a hardware change involved? Most ARMs have clock and voltages controlled from the kernel (e.g. I wrote OC patches for a Tegra2 kernel for the Toshiba AC100). Is the R-Pi similar? If so I'm not sure why this is news-worthy.
OC'ing won't kill the chip - just stop it booting -which is why there is an option to press shift during boot to disabled any clock change4s so you can set things back to a more sensible value.
Very VERY unlikely to overheat the chip either, unless you live in the Sahara, and even then onve it hits 85 all OC'ing is turned off anyway.
High overvolting could previously have set a warrantee bit (i.e. so Foundation can tell whether chip has been overvolted, and hence out of guarantee), now the level at which that bit is set is much higher, and better managed by the new drivers.
Newsworthy because its a load of kernel and other code ...
Yep. I haven't looked yet, but it seems that they'll have fixes for two problems I ran across...
1. No native kernel driver/firmware for some quite popular (read: cheap) wireless dongles.
2. Fix for excessive interrupt rate (dwc_otg.fiq_fix_enable=1 now the default) as mentioned in last paragraph.
I managed to find the fix for these myself thanks to the excellent forum and blog posts that people are making about the Pi, but it's certainly to be welcomed to have these baked in for less technically skilled users. As for the overclocking, the rpi version of xbmc has been overclocking to 800Ghz for quite a while (and there's been the option to do it in the /boot/config.txt in regular pi distros too). Nice to see that there's room to push the envelope even further and still be safe.
I remember there was once a website that sold CPU's that were guaranteed overclockable. reason being, they were hewn from a known batch etched on superior wafer. For this same reason, some cores are disabled in the factory (ie that particular run wasn't on top notch wafer)
I ordered a Pi from CPC today and THE BOOK from Amazon!