That should be a pic of a Preschott on fire at the top of page one. It'd certainly be more appropriate these days.

Thought about that....

...but I didn't think Mike would like a Prescott showing up on my expense account as, "destroyed for photo-op" But an 800Mhz Duron is almost disposable. (that's not photoshop, by the way, that's real flame on a real CPU. )

Whoa... I thought it was a photoshop. Did you use alcohol and burn that off? And was it still usable afterwards?

Tried alcohol and acetone, but the flame wasn't intense or long lasting enough. Actually, I experimented with this entirely too long....but I eventually snipped the heads off two matches, sat them on the die, and lit them while letting the camera snap a series of stills. Visually, the CPU is no worse for wear, but I haven't tried it in a system yet. (it had been bouncing around in the bottom of a parts bin for who knows how long, so I wouldn't know what to blame if it didn't work)



But besides the THG homage, any comments on the rest of the article?

"Toasty!!!"

And just for the record, besides comparing my CNPS7000A-Cu figures to past figures from Mike and Ralf, this is the primary method we used to calibrate my own P4 testbed.

So the mystery is off on what, "various methods," were used to correct the readings on my test bed, as mentioned in the reviews.

Very nice. I'll have to return to it when I have the time. I don't think I should worry about my CPU temps when my HDD easily hits 45°C while defragmenting, MBM recorded 50°C as the highest it reached Idle stays consistent at 31-32°C though.

I agree, it looks photoshopped , show us some aftermath pics!

Thanks for the arcticle, Rusty. It takes a while to run all the tests, so feedback may be slow. CPU temperatures are a current interest of mine since on my s754 board temps average 70C folding with a DH7-CG, while the SH7-CO temps are around 50C. My temps are down from @ 80C with the xp120.
I tested with prime95 small fft's as this makes the most consistant temps with the temps (graphed w/ speedfan) remaining btw 2C and often 1C. I used the first plateu, about 10min. The temps would rise a degree or two on extended runs. All settings were default, using multiplier changes to increace cpu speed, bus speed changes for the 2 highest speeds, 800M to 2.4Ghz at 200Mhz intervals.
My results were linear and the correction was -26C!!!
I have a couple of problems with my result that I hope yopu can help me with. First, this results in a corrected idle temp of ambiant ( 5C lower than a temp probe tip touching side of case). Is this normal and if so why? Also, how is the increase in effectiveness of the hsf as delta-T rises compensated for? I'm guessing the exponential rise in wattage. A third problem is that my idle temps barely changed at all from 51.5 to 52.5 over an increace of cpu speed of 1.6Ghz ( .5 is the graph zig-zagging inside a 1 degree seperation).
I have a predjudice that cpu temps are commonly much higher than reported. Any help that gives me a better understanding would be appreciated.

woah. That's pretty, um, impressive (?) Just to be sure, did you double check your MBM settings and your math?





Given the Mhz and Vcore drop of the Cool-and-Quiet, I would expect the idle to be pretty low, but below ambient does seem to be violating some laws of physics. A couple of ideas: What is the CPU cooling system? Is it drawing in air that may be cooler than you think it is? Try putting you temp probe in the intake stream of the HSF. Or...perhaps the offset becomes non-linear once the extreme drop in CPU wattage occurs when the C&Q kicks in.

Since you're the first A64 guinea pig, I'm not really sure how typical your results will be. (plus the fact that your temps have been goofy from the very beginning. )






I've had similar thoughts. My current theory is that the range is temps we're seeing is so small that the increased efficiency of the heatsink is lost in the relatively poor resolution of the temp probe. If we were using temp reporting equipement that had higher than 1° of resolution we would likely see a drop in °C/W as the temps rose. Comparitively speaking, these are still pretty imprecise temperature readings.

Any chance Cool n' Quiet could be affecting these results? Have you tried running the calibration test array with CnQ disabled? It would be interesting to see if CnQ affects this procedure in any way.

BTW I have also heard many reports of the Mobile or DTR chips have temperature report issues on certain boards; for example, once you flash the new DFI LANParty UT NF3 350g with Oskar Wu's beta BIOS to resolve the Vcore-on-boot issues, it supposedly screws up the diode reports from the DTR or Mobile chip.

Anyhow, I'll get a chance to try out the procedure on my own K8 as soon as my new motherboard comes in, and then we'll see if my K8 does the same. I'll test with this XP-90 I've got here, since the water cooling setup would take too long.

Anyhow; it will be nice to have a calibrated board to begin with, assuming the procedure works right.

-Ed

Thanks Ed, All vestiges of C'nQ have been removed from my computer. On my board it doesn't agree with overclocking or manual voltage settings. The temp diff was a degree or 2 at idle.

I resolved algebraically, the result was the same. -26C HS=2400 LS=800 HT=70.5 LT=58

Room temp was 74f, not much cooler, and my fan is in suck mode, pulling air off the board. My ambiant is the air temp entering the psu. Case air pressure would be negative but I have appx 1sq ft vent openings.

Any bios for s754 that attemps to compensate for the SH7-CO and DH7-CG cpu's will have simular problems. The SH7-CO Clawhammer with disabled cache sold as Newcastle (oem# ending -xxx4AP) and the DH7-CG a small die native 512k Newcastle (oem# ending -xxx4AX) as the diode report @ 20C different with the same bios. My belief is that the Clawhammer -xxx5AP and the revised Clawhammer xxx4 or 5 AR Newcastles report somewhere between but I have limited data.

Haven't gotten to run CPUID on mine yet, but I do have it and it's an AX (CG) NewCastle, for your reference.

According to eMicroX, my board ships today, via 2-day, so I should have the board Monday.

I think I'll leave the IHS on for now, with the XP-90 to calibrate, and then once I'm done with that, switch to my Storm G4 block and remove the IHS.

Stew Forster specifically optimized the block to overclock better, by more evenly cooling the entire die, rather than just build a block that reads lower temps on the diode. His statement is in this reply in the Storm internal pics thread. As such, I see no reason whatsoever to keep the IHS on it, as I personally am not prone to crushing or chipping bare cores.

-Ed

Until we get a few more reports from A64 owners, I'm not sure what you tell you Glassman... hopefully more reports will help.


Glassman's nice explanation of the temperature reporting woes of the A64, combined with the whole A64 MDP issue, is why you're not going to see any SPCR A64 HSF reviews in the near future. Just too many unknown variables to make any reports meaningful.

Just check the heat spreader to confirm. -4AX is all you need to confirm the real Newcastle.
Here's the link http://www.amd.com/us-en/assets/content ... /30430.pdf
for all the amd specs. Removing the ihs should give you a C/w of @.6 instead if .3 for a lot smaller area. (PS your link was to Cathar's new block) Couldn't find the IHS ripped off (Are you crazy?)
Here is a link to my first ever thread, which is about the temp diff. I'm very proud it went 3 pages. http://forums.pcper.com/showthread.php?t=353817 This thread disagrees with my theory of internal temps, so I hope to learn more about the subject from SPCR members.

Just a quick note: For readers' convenience, the xls file used by Russ to calculate for c has been post on the 2nd page of the review.

The chip is marked AX, as I said, but since you brought up the 4, I just checked and yes, it's 4AX.

Yes, my link is supposed to be to a post on Cathar's block; it's his statement regarding the manner in which he designed the block to remove heat from a bare die, and explains the reasoning why I would prefer to have direct-to-die contact with the block. There are cases where the IHS functions as desired (which is why I will run the calibration procedure with the XP-90 and the IHS not removed), helping to spread head from the hotter parts of the die to the cooler parts to more efficiently cool the die as a whole; however, in this (the Storm G4's) case, it is clear to me that Stew's block does a better job of this than AMD's IHS will; either way, it is generally a fact that the more boundaries you have between two surfaces, welded, TIMmed, or otherwise, you will lose efficiency, and this is just another layer of loss being bypassed.

Btw, there are quite a few desktop K8 users who have removed their IHS; I've not yet reached the stage of soldering additional components to my mainboard or using phase change cooling, so I don't quite think I've reached crazy (yet).

-Ed

Quick and dirty with little error checking

Sorry, Ed misread your post, thought it was coming with the board. 4 is the cache size, 512k. 3=256, and 5=1M. I'll be interested in your delta t's. I'm wondering if the changes will effect the constant. Will wait and see.

Good job, Russ. However, the article really should include a warning about variable-speed fans. Constant HSF RPM is a given, but even PSU fan speed could affect results in a case.

Dang, good thinking.....a variable speed CPU fan could seriously muck things up.

The PSU fan would very little effect, if you're following the directions and measuring the ambient temp at the HSF's intake.

I guess that's what we get for always testing in an open test bed!

OK, some more A64 results.

First the test details: The motherboard is Abit KV8Pro. I disabled both CnQ and Abit EQ (messes with fan voltages), just left the "turn off PC if CPU is 85C" on.

I had to do the tests twice. The first time around the case door was open as the PSU I was using was a temp one (the installed-with-much-time-and-effort Antec 380s was unstable) sitting outside the box. Plus, I thought I could just change the multiplier and run the tests again without waiting for the CPU to cool properly. Wrong. It messed up the results. The second time round, I reconnected the Antec PSU (it takes some time for it to become unstable), closed the case door and properly cooled the CPU in between.

The PSU is modded with an acoustifan with the thermistor intact. Plus, the PSU has its own voltage adjustments. However, I used the rpm line to make sure that the fan rpms also stablized before taking any readings.

I used a freshly installed Windows 64bit edition. I was planning on using the BurnIn64 - a 64bit load program that does some fast fourier transforms but tests revealed that the K7 CPU Burn program loaded the CPU more and faster. On the average, it took 30 minutes for the CPU to reach steady state. I used ClockGen to lower VCore to allow faster cooling but readjusted VCore and waited at least 5 minutes before attempting another burn.

I used my own excel table (Russ's hadn't been posted yet), included the VCore readings from MBM as well. I used Excel's "add trendline" function to get the best fit.

I used a FSB of 200 (mobo doesn't allow lower) and used clockgen to change multiplier between 10 and 5 in 0.5 increments. Those are the limits of the BIOS. I wanted to try faster FSB but at 2000 MHz @ 1.5V the CPU hit 84C, 1C less than shutdown temperature so I didn't try.

When plotted, the points form a very nice line. Intercept point was 3.205C. Excel says there's a good fit with R2 = 0,9979.

Formula: y = 0.0221x + 3.205 where x is frequency, y is delta temp

Unfortunately, in the university I had a VERY picky professor who introduced us into technical drawing and I've become perceptive of small curvatures. And this line had a very slight curvature. So I tried a polynomial trendline to see what I would get. The polynomial trendline fit almost perfectly with an R2 value of 0,9998. The x2 factor was in the 1/million range which is why it looked straigt. With the new trendline, intercept became 11.958.

Formula: y = 0.000004x2 + 0.0105x + 11.958

So now, what is my calibration? +3.2C or +12C?

/Surgeon general's warning: Too much math mugs your brain /

burcakb --

Wow!
Interesting post.

But you don't tell us what temps you're actually getting. Also [* warning! Big conjectures coming *]...

The assumption of this calibration is that there is a direct relationship between CPU power/temp and CPU clock speed. In general, it's true, but there are always non-linearities in any machine process (or even natural phenomenon*). Also, if you are dealing with a CPU that has a heat spreader, I believe the thermal non-linearities will be greater. So I'd probably pick the number that gives more consistent results across the board & shurg about the one or two points that don't make sense.

------------
* Totally Off-Topic Digression:

AFAIK, tmost striking exception to natural laws (on earth) discovered by science is the behavior of water as temp varies. Like most things, water expands as temp is raised. Unlike most things, water expands when it freezes. This is why ice floats, its density is lower than water; and why beer bottle break when stuck in the freezer for a quick cooling and then forgotten (the water in it expands). Without this characteristic of water, there would be no life on earth: Glaciers would break out into icebergs that sink to the bottom of the ocean, where the lack of heat would ensure they stay forzen forever. The ocean temp would drop from bottom up and all the oceans would eventually be frozen. No oceans, no life.
------------------
ok - end of digression & fantastic conjectures & your turns to take batting practice at Mike's expense.

Very interesting Nice work!

Could you email me your excel file? I've got a couple of theories, and would like to see the data to try them out.

Mike,

Max temp at 2000 Mhz was 84C, max temp at 1000 MHz was 63C. VCore was around 1.51V-1.52V so that was pretty constant.

Case temp reports and room ambient stayed rock solid stable throughout the effort. One interesting thing was that C/W values started going up as clock speed went down. Depending on how you scale the thing, it looks straight or not. The fanmate2 may have been to blame for that (even though I didn't touch it) because uGuru somehow can't get reliable fan rpm data through it.

At some point I might think about retrying it at a lower VCore than default (I regularly run my machine at 1.25V) but it's already cost me two days of folding.

Rusty, mail coming up...

Off-topic!

About the water-ice anomaly posted by MikeC:

For those of you who don't know and are interested, there's really nothing strange about water's density being lower when frozen. It's because of the water molecule's geometry - it looks like a "V" with a positive and a negative end. When it get's cold enough (<4 degrees C), because of the molecule's differently charged ends, it starts to form a crystalline structure composed of "rings" (well, almost). This leaves a lot of space in the center of the rings, which results in the structure of water molecules in ice take up more space than the more tightly packed molecules in liquid water.

http://www.nyu.edu/pages/mathmol/modules/water/info_wat.html has a little more reading for the interested.

Burcakb, I solved for your numbers, and assuming an ambiant of 30C (cpu intake) I came up with the result of -12C. Multiply by 1.1 if top temp refers to 1.9G. Your corrected temp would be 72C. ( A degree more ambiant = 1 less degree correction)
Speaking as some one who ran without issue at 79C (-26=53C) I'm concerned that you are running awfully hot at stock speed and voltage. If your fan speed doesn't account for your temps, check your hsf seating, as I only saw temps like yours when I was running 2.5G @1.7v, and my correction is much larger.

While perfect fits are good for technical drawing, they probably aren't as good for modelling. Adding the polynomial term requires a degree of freedom and decreases the number available for error. Since you're starting from 11 points, I suspect the increase in fit isn't enough to warrant the additional of the quadratic to the model. The real test would a significance test for your quadratic coefficient, and given the closeness to zero, I suspect it's not significant.

Your interpretation is off too. What you're doing with those equations is predicting temperature (or delta temp) based on speed. Your intercept there is the estimated temperature at speed zero -- which should be your ambient temps (or zero). In this context, the linear intercept also makes much more sense.

But going back to the original article numbers, to get your temps of 63C and 84C to match your equations, you need to offset them by 35 - 36.5 C. Plugging those numbers and temps into the article equations gives an offset of -39 (which is again very different than the intercepts). Those number makes your temps sound a little more reasonable.

But back to your ambient temps -- that 35-36.5 C sounds like you're running hot because of inadequate case cooling.