Re: Mersenne: mprime for QA or performance?
Hi, At 12:24 AM 1/13/99 -0800, Leu Enterprises Unlimited wrote: On one CPU, after two days worth of burn-in, the time required to complete 100 iterations on the stock number went down, as I would expect. On a second CPU, the time actually *increased*. From 1 day, 18 hours, and 55 minutes, to 1d 19h 13m! I'd be surprised if burning in had an effect on iteration times. More likely, 100 iterations is not enough to get a truly accurate timing. I don't know about Linux, but it has been noted before prime95 iteration times can vary a few percent from day to day. If anyone knowledgeable would care to comment about mprime's suitability for Q.A. mprime is great for QA. It generates heat (by FPU use) and tons of memory accesses. If there are any hardware problems there, they are likely to show up in an extended torture test. and/or performance measurements, I'd guess its as good as but no better than any of hundreds of publicly available benchmarks. Best regards, George
Re: Mersenne: mprime for QA or performance?
On Fri, 15 Jan 1999, George Woltman wrote: At 12:24 AM 1/13/99 -0800, Leu Enterprises Unlimited wrote: On one CPU, after two days worth of burn-in, the time required to complete 100 iterations on the stock number went down, as I would expect. On a second CPU, the time actually *increased*. From 1 day, 18 hours, and 55 minutes, to 1d 19h 13m! I'd be surprised if burning in had an effect on iteration times. More likely, 100 iterations is not enough to get a truly accurate timing. I'd be extremely surprised, since the speed of the computations is strongly linked to the processor clock, so either you're measuring clock drift (and yes, that is temperature linked) or due to your small samples the variance mentioned which is about 1% is actually noice. -- Henrik Olsen, Dawn Solutions I/S URL=http://www.iaeste.dk/~henrik/ Get the rest there.
Re: Mersenne: mprime for QA or performance?
To me at least, the first part is obvious. As temperature goes up, so does resistance. The higher resistance requires a greater potential difference to send an adequeate(sp?) amount of current throught the processor. Without cooling, I would imagine the chip would eventually melt down. With cooling, I would imagine the voltage peaks out in some sort of gaussian curve. For small temperature ranges then R = kT + c, the trouble is, for semiconductors, k can be negative. This is why semiconductors can ( do) burn out if overrun or undercooled - current can "run away" to very large values (essentially whatever the supply rails will stand before *they* step in act as fuse wire) When running digital chips very fast, what you need is a sharp rise/fall. Now you just *never* see nice square waves like those drawn in the text books, real signals tend to look like a sine wave. Obviously, in these circumstances, the higher the voltage, the steeper the rise fall of the signal when it's changing from a 0 state to a 1 state, or vice versa. It's also why small die sizes can be run faster than large die sizes, the current is much smaller, so it takes less power to change from 0 to 1 in a given time. Which brings to mind a theory about the second part, "speed increase". Too much heat will damage a processor. However I wonder if there is a relationship between some sort of particle drift (electron???) and resistance that "breaks in" or "burns in" a new processor? Much like the performance on a new engine after it has been carefully broken in after the first 1,000 miles... I get the impression that failure rates on (non-overclocked) semiconductor device falls rapidly with time, at least over the first few hours. This is why quality systems suppliers "burn in" or "soak test" systems for 24-48 hours before shipping. It may be the case that minor flaws can act as "pinch points" which eventually cause burn out failure, even though the chip actually measures within spec to begin with. Also atoms within crystals can "creep" causing minute changes to the circuit characteristics, I would have thought that the most significant effect would be the first time the circuit gets to its "normal" operating temperature. Regards Brian Beesley
Re: Mersenne: mprime for QA or performance?
To my knowledge, the effect on performance of burn-in has not really been widely discussed or explored; at least not on any of the main overclocking forums which I am aware of. Which is partly why I raised the question. Some things are known to happen; and I have observed them myself. First, the amount of voltage required to work at a given speed decreases. Secondly, the odds of the chip working at a higher speed increases. This much is common knowledge. No one has satisfactorily explained these results, though many theories abound. There are also rumors that the higher-quality chip cores, which get put into the C333A/366/400 and Pentium II's are measureably faster. These rumors have not been confirmed to my knowledge. What I was saying was that an improvement in speed was not unexpected; I.e. it is consistant with the expected improvement in the chip due to burn-in. These are also my own expectations. It would be difficult to classify unexplored territory as "common knowledge". Perhaps it's common knowledge to some; it's just not mentioned in the more popular websites or forums. -dwight- From [EMAIL PROTECTED] Wed Jan 13 18:17:14 1999 This brings up an interesting question (Well at least to me it is). I notice that you refer to the iteration time "going down" as an expected event. Are you saying that a new CPU is expect to get faster (If ever so slightly) after an initial "Burn in" time??? Perhaps that is a bit of common knowledge I have never heard of. Can someone explain this in greater detail? Thank You, Chuck On Wed, 13 Jan 1999, Leu Enterprises Unlimited wrote: Greetings! Aside from it's normal uses, I've been looking at using mprime as a QA and benchmarking tool. This is particularly important for overclocked Beowulf clusters, where reliability is a must. And given that the price of these is now so cheap (with a Gigaflop being achieved for under $10,000), adaquate Q.A. procedures are a must. Re: my observations on www.supercomputer.org. I've recently noticed something interesting. Namely, with the so-called effects of "burning-in" a CPU via mprime's torture test, and the resulting effects on the CPU speed (as determined by the mprime time test). On one CPU, after two days worth of burn-in, the time required to complete 100 iterations on the stock number went down, as I would expect. On a second CPU, the time actually *increased*. From 1 day, 18 hours, and 55 minutes, to 1d 19h 13m! So this leads me to wonder if what I'm seeing is real, or whether mprime is really sensitive enough to be used here. The O.S. being used is Linux. If mprime is indeed seeing some real effects here, I think people need to be aware of it. Or even if it's not suitable for this, people should also know. If anyone knowledgeable would care to comment about mprime's suitability for Q.A. and/or performance measurements, I would appreciate it. Thanks! -dwight-
Re: Mersenne: mprime for QA or performance?
Some things are known to happen; and I have observed them myself. First, the amount of voltage required to work at a given speed decreases. Secondly, the odds of the chip working at a higher speed increases. To me at least, the first part is obvious. As temperature goes up, so does resistance. The higher resistance requires a greater potential difference to send an adequeate(sp?) amount of current throught the processor. Without cooling, I would imagine the chip would eventually melt down. With cooling, I would imagine the voltage peaks out in some sort of gaussian curve. Which beings to mind, would the potential difference delta be a good measure of the quality of your cooling? (For the non electrical people, I am using voltage and "potential difference" interchangeably, they are the same thing but potential difference says it better IMHO) Which brings to mind a theory about the second part, "speed increase". Too much heat will damage a processor. However I wonder if there is a relationship between some sort of particle drift (electron???) and resistance that "breaks in" or "burns in" a new processor? Much like the performance on a new engine after it has been carefully broken in after the first 1,000 miles... This much is common knowledge. No one has satisfactorily explained these results, though many theories abound. Case in point... There are also rumors that the higher-quality chip cores, which get put into the C333A/366/400 and Pentium II's are measureably faster. These rumors have not been confirmed to my knowledge. Tighter packed transistors = less ability to dissipate heat = more resistance... It would be difficult to classify unexplored territory as "common knowledge". Perhaps it's common knowledge to some; it's just not mentioned in the more popular websites or forums. Almost everything that I have learned beyond 1st grade was by assuming that it was common knowledge. "If someone else understands it, I should be capable of understanding it too!"... ~~~ : WWW: http://www.silverlink.net/poke : Boycott Microsot: : E-Mail: [EMAIL PROTECTED] : http://www.vcnet.com/bms: ~~~
Re: Mersenne: mprime for QA or performance?
From [EMAIL PROTECTED] Thu Jan 14 13:25:19 1999 To: Leu Enterprises Unlimited [EMAIL PROTECTED] cc: [EMAIL PROTECTED] Subject: Re: Mersenne: mprime for QA or performance? Some things are known to happen; and I have observed them myself. First, the amount of voltage required to work at a given speed decreases. Secondly, the odds of the chip working at a higher speed increases. To me at least, the first part is obvious. As temperature goes up, so does resistance. The higher resistance requires a greater potential difference to send an adequeate(sp?) amount of current throught the processor. Without cooling, I would imagine the chip would eventually melt down. With cooling, I would imagine the voltage peaks out in some sort of gaussian curve. Which beings to mind, would the potential difference delta be a good measure of the quality of your cooling? That's not been my experience. You can put on superb cooling, and crank the voltage up all you want. It won't help a bad chip. You will only fry the chip. Don't forget that the CPU parts are spec'd at 2.5V; with a typical tolerance of +/- 20 %. Running a C300A at 3.0+ V for long is likely to cause great disappointment. Furthermore, you are not guaranteed to run reliably at 504 MHz, if your CPU handles 450 MHz @ 2.0 V. I have several such CPU's. And, FWIW, the point of failure appears to be the FPU (no - not the L2 cache, like most people assume). ... It would be difficult to classify unexplored territory as "common knowledge". Perhaps it's common knowledge to some; it's just not mentioned in the more popular websites or forums. Almost everything that I have learned beyond 1st grade was by assuming that it was common knowledge. "If someone else understands it, I should be capable of understanding it too!"... Ok - please explain to me members of the opposite sex. They certainly seem to understand each other. :) Seriously, I wouldn't classify the original subject as common knowledge among overclockers at all. Or the latter subject, but let's not digress... Back to my orginal point - is there truely no one on this list who can knowledgeably state that the precision of mprime is suitable, or not suitable, for these timing purposes? One area of concern is round-off error. But there may be others as well. TIA! -dwight- To learn how to build your own supercomputer, for under $10,000, go to: www.supercomputer.org
Mersenne: mprime for QA or performance?
Greetings! Aside from it's normal uses, I've been looking at using mprime as a QA and benchmarking tool. This is particularly important for overclocked Beowulf clusters, where reliability is a must. And given that the price of these is now so cheap (with a Gigaflop being achieved for under $10,000), adaquate Q.A. procedures are a must. Re: my observations on www.supercomputer.org. I've recently noticed something interesting. Namely, with the so-called effects of "burning-in" a CPU via mprime's torture test, and the resulting effects on the CPU speed (as determined by the mprime time test). On one CPU, after two days worth of burn-in, the time required to complete 100 iterations on the stock number went down, as I would expect. On a second CPU, the time actually *increased*. From 1 day, 18 hours, and 55 minutes, to 1d 19h 13m! So this leads me to wonder if what I'm seeing is real, or whether mprime is really sensitive enough to be used here. The O.S. being used is Linux. If mprime is indeed seeing some real effects here, I think people need to be aware of it. Or even if it's not suitable for this, people should also know. If anyone knowledgeable would care to comment about mprime's suitability for Q.A. and/or performance measurements, I would appreciate it. Thanks! -dwight-
Re: Mersenne: mprime for QA or performance?
This brings up an interesting question (Well at least to me it is). I notice that you refer to the iteration time "going down" as an expected event. Are you saying that a new CPU is expect to get faster (If ever so slightly) after an initial "Burn in" time??? Perhaps that is a bit of common knowledge I have never heard of. Can someone explain this in greater detail? Thank You, Chuck On Wed, 13 Jan 1999, Leu Enterprises Unlimited wrote: Greetings! Aside from it's normal uses, I've been looking at using mprime as a QA and benchmarking tool. This is particularly important for overclocked Beowulf clusters, where reliability is a must. And given that the price of these is now so cheap (with a Gigaflop being achieved for under $10,000), adaquate Q.A. procedures are a must. Re: my observations on www.supercomputer.org. I've recently noticed something interesting. Namely, with the so-called effects of "burning-in" a CPU via mprime's torture test, and the resulting effects on the CPU speed (as determined by the mprime time test). On one CPU, after two days worth of burn-in, the time required to complete 100 iterations on the stock number went down, as I would expect. On a second CPU, the time actually *increased*. From 1 day, 18 hours, and 55 minutes, to 1d 19h 13m! So this leads me to wonder if what I'm seeing is real, or whether mprime is really sensitive enough to be used here. The O.S. being used is Linux. If mprime is indeed seeing some real effects here, I think people need to be aware of it. Or even if it's not suitable for this, people should also know. If anyone knowledgeable would care to comment about mprime's suitability for Q.A. and/or performance measurements, I would appreciate it. Thanks! -dwight- ~~~ : WWW: http://www.silverlink.net/poke : Boycott Microsot: : E-Mail: [EMAIL PROTECTED] : http://www.vcnet.com/bms: ~~~
