My Big Bad Building Guide/Lapping Guide/Overclocking thread

[DrMrLordX]

**DISCLAMER: This post is not a specific endorsement of any particular product or manufacturer/reseller. I just used parts that I felt would work well within my budgetary constraints given what I wanted the system to do. Feel free to use whatever works well for you. Investigating products from Anandtech's various advertisers is strongly encouraged**

The upgrade bug bit me a few months back, so I've been looking for excuses to buy new hardware and tinker with it. Here's what I bought:

Case: Coolermaster Centurion 5
PSU: Corsair HX520
Motherboard: Abit NF-M2 nView
CPU: AMD Athlon X2-3600+ (Brisbane) OEM
RAM: OCZ PC2-6400 Platinum DFI Special (2x1gb)
HSF: Thermaltake Big Typhoon VX
HDD: Seagate 7200.10 250gb (16 mb cache) OEM
DVD burner: NEC AD-7170A OEM
Extra fan: Silverstone FM-121
Thermal Paste 1: Arctic Silver Ceramique (ordered AS5, got Ceramique)
Thermal Paste 2: Shin-Etsu X23-7783D

Plus some zip ties, zip tie mounts, electric tape, duct tape, a piece of window glass, 200/400/600 grit silicon carbide wet/dry sandpaper, two Sharpies, and . . . well that's about it I guess.

Q: No video card?
A: Yes, no video card. That might come later, but for now I'm using the motherboard's onboard video.

Q: No sound card?
A: uh, yeah.

The result is: a working computer! I even filmed the whole ordeal and spliced together a building guide aimed an newbies that might be a bit intimidated by the idea of building a PC or haven't otherwise seen one built before. It's far from comprehensive, but it covers most of the important bases, again minus a video card because, hey, I didn't really need one. Much. Yet. Here's the video. If you like it or think it might help somebody out, feel free to tell others about it. A few build notes:

1). Cable management (or lack thereof):
I went into the build process armed with zip ties, zip tie mounts, and even electrical tape thinking I could engage in some fancy cable management. Then I realized that the case had so few holes in the motherboard tray and elsewhere in the case that it would have been very difficult for me to do anything fancy like this. What I did instead was move most of the few power cables I did need to use into the empty 5 1/4" drive bays where air didn't need to move (in this setup anyway) and restrict some other cables from falling into fans or otherwise interrupting airflow with zipties. It worked out very well. Airflow in the case is good.

2). Why I chose what I chose:

Coolermaster Centurion 5: I knew I wanted a case with a side air duct/fan/hole/vent/etc. so I could use the Big Typhoon (in this case, VX) with a Silverstone FM-121 mounted on it and get unrestricted airflow for the CPU without having to buy a high-airflow case like the Antec 900. At around $55, the price was right and it fit the bill. The Big Typhoon gets all the cool air it wants and keeps most of the motherboard cool as well.

Corsair HX520: 80% efficiency, solid dual 12v rails, reliable components, and modular cables. Yeah, that works for me. It leaves plenty of room for upgrades including a single g80/r600 video card.

Abit NF-M2 nView: Okay, so it's AM2. See the next entry on why I went AM2. That aside, I had it down to two boards: this one or the DFI NF UltraII-M2. The Abit board was cheaper and had onboard video so I could get it and avoid buying a $20-$40 el-cheapo vid card that I would openly resent. I do sort of miss having the OC options of the DFI board, though, but I'll live.

AMD Athlon X2-3600+ Brisbane: Several reasons. First off, I've been curious about the impact of the DDR2 memory controller on AM2 chips for awhile, but nobody has really done a lot of memory benching here on Anandtech with the AM2 platform (that I've noticed). Secondly, I like an underdog. Thirdly, I arrogantly believed that I could beat the 3.1 ghz wall so many other people hit by using a combination of good thermal paste, good HSF/fan combo, and lapping. Fourthly, I wanted to see how well X2s scale vs Core 2 chips at high clock speeds (based on this benchmark and testing I hope to complete on my own, see the overclocking post below). Fifthly, the price was right. ~$100 for an X2? Sweet. EDIT: now $55 for an X2? Unbelievable.

OCZ PC2-6400 Platinum DFI Special: The kit works well, the board supports its upper vdimm limit of 2.4v, and it was $160 after mail-in rebate. This should give me plenty of wiggle-room when testing the Brisbane's memory controller and its effects on scaling.

Thermaltake Big Typhoon VX: I've seen at least one benchmark where the Big Typhoon dominates provided it has a 100+ cfm fan. The VX ships with an 86.5 cfm fan, but I unmounted that and used it as the exhaust fan for the Centurion 5. I kind of wish I had gone with the new Ultra 120+ instead, but it's not even out yet, and I doubt it would cool the motherboard components as well as the Big Typhoon does. Plus, the VX was not that expensive. I had to get the VX to be sure I got the right mounting mechanism for AM2 as well.

Seagate 7200.10 250gb (16 mb cache): It's the cheapest 7200.10 drive with 16 mb cache. Given the fact that the 320gb 7200.10 handily defeats the 750gb 7200.10 performance-wise, I figured I'd try out the 250 gb drive and see if it was as good as the 320 gb one. So far, so good, but I haven't done any testing.

NEC AD-7170A: Eh, it works. I liked my 3520A so I figured I'd give NEC's latest a whirl. This isn't just an NEC drive, it's a Sony NEC Optiarc drive. It seems okay to me.

Silverstone FM-121: After removing the VX's included 86.5 cfm fan and using it as the case exhaust fan, I strapped this 110cfm beauty onto the Big Typhoon for extra cooling action. Since it's only a 120x120x25mm fan (vs 38mm), it mounts with ease.

Arctic Silver Ceramique: Originally I ordered AS5 with my Big Typhoon VX as a knee-jerk reaction. Then I did some reading and found out that Shin-Etsu x23 kicks its butt all over the place. So I 2-day shipped some X23 so it would get here at the same time as the last of my parts. When my last shipment came in, I discovered that the shipper sent me AS Ceramique instead of AS5. Interesting, but hardly noteworthy since I wound up using X23 instead.

Shin-Etsu X23-7783D: As mentioned above, this stuff rules, especially on lapped surfaces. And I was gonna lap, so I got some. It's hard to find, but if you Google search long enough (or just read the thread I linked above) you can find it. 7783D is the latest variant.

3). Build quirks:

Case faceplate: For some reason, odd plastic bits jut out in the 5 1/4" drive slots preventing your CD/DVD drive or other 5 1/4" drives from sticking all the way out to be flush with the faceplate. I think. Maybe there's something I'm missing here? Anyway, right now, my NEC drive is about 1-2" recessed into the machine, though the tray still opens without difficulty. It works fine, it just looks weird.

Narrow 3 1/2" drive bay: Dunno why, but the harddrive cage is almost too narrow. It scratched up the sides of my Seagate drive. The locking mechanism doesn't seem to work properly either, or at least, not terribly well. Still, the drive doesn't slide around because the bay is so narrow. Very snug fit.

Bizarre temp displays: The NF-M2 nView and Brisbane together produce wild and crazy CPU temperature readings. The motherboard temp sensor reports temps around 30C too high at least, and there's no telling if it moves upward in temperature properly when the CPU is at load. Core Temp, on the other hand, reports impossibly low core temperatures. All the other sensors on the board seem to be working very well. The only way I have to check CPU temperature right now is to put my hand on the base of the heatsink, check the sides for heat leaking out from the socket, and uh, guess. Things seem to be pretty cool right now.

Cranky CPU fan header: The FM-121 could not draw power when connected to the 3-pin CPU fan header on the motherboard. The motherboard also complained audibly that I had no CPU fan installed when I did (even though it would not power up). The first 30 seconds this CPU spent in operation were without a fan blowing on it at all save the exhaust fan which was only pulling air at 1200 rpm at the time (low setting). I had to connect the FM-121 to a 4-pin molex and deactivate the CPU fan alarm in the BIOS to avoid obnoxious alarm beeps.

And . . . that's all! Hope you enjoyed the video. The next two posts will have lapping and overclocking information.

 

[DrMrLordX]

I've read a lot about lapping CPUs and HSFs, but I had never done it before this build. Now I can honestly say I've done it once. Sadly, I lapped as I built, so I have no temperature comparatives. In fact, I'm not even entirely sure that it helped or that it was necessary. Everything seems to be working very well though, and I can tell you that the CPU IHS was NOT flat. I filmed parts of the lapping nonsense for your amusement. Boy, were my shoulders sore after all this lapping nonsense.

I'm still not 100% sure that I did it right, but based on what little I know about my CPU and core temps, it seems to have gone well. I am not getting any obvious symptoms of poor contact between the HSF and IHS.

Also, while many go for mirror finishes while lapping, I stopped at 600 grit paper.

Most of my other notes about the lapping are in the video, so . . . have at it.

 

[DrMrLordX]

Overclocking is still in progress. Some random notes:

1). vID on this chip is 1.3v. Current speeds by voltage:

1.3v: ~2850 mhz
1.45v: ~3080 mhz
1.475v: ~3134 mhz
1.525v: ~3200 mhz

At this point, getting stable clocks above 3.1 ghz seems damn near impossible. Takes too many volts. I may have to reseat the HSF and make sure my lapping job was good . . .

2). Memory clocks are strange on this board, or maybe it's the chip? The 1:1 ratio produces memory clocks that you would expect, but the other ratios produce lower memory clocks than they should. The lower the CPU multiplier, the larger the speed penalty on the RAM. I need to do more testing, but it seems like instead of the usual formula of:

(CPU speed/(CPU multi/DRAM divider)) = DRAM speed

I get

(CPU speed/(CPU multi/DRAM divider)) * (CPU multi/10) = DRAM speed

In other words, take the RAM speed you'd expect, divide your CPU multi by 10, and then multiply it by the expected RAM speed. For example, with an HTT of 200, a 9.5 CPU multiplier, and the 1:2 memory ratio (DDR2-800 as its called in the BIOS) I get a RAM speed of 190 mhz, or DDR2-760. If I drop the CPU multi down to 8x, the RAM speed drops to 160 mhz (DDR2-640). I wonder if this isn't some part of the power-saving enhancements on Brisbane? Maybe when cool n' quiet kicks in, it underclocks the memory as well?

3). On this board I've gotten the RAM to run at DDR2-1100, 5-5-5-15, 2.4v vdimm with very little tweaking. I could give it 2.5v later if I feel like voiding the warranty.

More coming later, and I'll probably film this as well if the camera cooperates.

 

[DrMrLordX]

I finally got around to some of the testing I wanted to do . . .

first off, I wanted to see how well an E4300 would scale performance-wise at higher clock speeds. I wound up using an E6300 to guess at the scaling of low-end Conroe processors due to the fact that an E6300 at high clock speeds should scale better than an equivalently-clocked E4300 due to higher FSB speed, higher memory speed, etc (provided also that memory timings are the same). Of course, memory strap issues complicate things a bit there but . . . overall, the E6300 should still scale better.

Anyway, from the data available on the above link, I examined Super Pi results and came to the conclusion that the E6300 scaled at approximately 77% efficiency. Namely:

The reduction in pi time from 1860 mhz to 3500 mhz was 9.838 seconds
The reduction in pi time should have been ~12.723 seconds had the processor scaled at 100% efficiency.

The processor only achieved 77% of the estimated performance.

Then I recorded the following super pi data from my X2-3600+ Brisbane:

1800 mhz (200x9) DDR2-600 5-5-5-15 2T
result: 50.531s

2253.46 mhz (250.38x9) DDR2-751.16 5-5-5-15 2T
result: 40.281s

2701.74 mhz (300.19x9) DDR2-900.58 5-5-5-15 2T
result: 33.469s

3060.11 mhz (340.01x9) DDR2-1020.04 5-5-5-15 2T
result: 29.562s

3104.12 mhz (344.90x9) DDR2-1034.70 5-5-5-15 2T
result: 29.125s

3148.79 mhz (349.87x9) DDR2-1049.60 5-5-5-15 2T
result: 28.672s

3194.75 mhz (354.97x9) DDR2-1064.92 5-5-5-15 2T
result: 28.297s Checksum: 8C56A504

First off, let me be absolutely clear that Brisbanes are not going to overtake Core 2 chips in Super Pi! Scaling aside, the performance of Brisbane at 1800mhz was appallingly bad. Granted, the memory was kinda slow compared to the E6300 sample from the article I linked, but still . . . 50 seconds? Good grief.

However, I processed the data points and came to the following conclusions:

1800 mhz-> 2253.46 mhz
25.2% increase in clock speed
Estimated Super Pi time at 2253.46 mhz given 100% scaling efficiency: 44.166s
Actual Super Pi time at 2253.46 mhz: 40.281s
Scaling efficiency: ~109-110%

1800 mhz -> 2701.74 mhz
50.1% increase in clock speed
Estimated Super Pi time at 2701.74 mhz given 100% scaling efficiency: 37.874s
Actual Super Pi time at 2701.74 mhz: 33.469s
Scaling efficiency: ~113%

1800 mhz -> 3194.75 mhz
77.5% increase in clock speed
Estimated Super Pi time at 3194.75 mhz given 100% scaling efficiency: 30.953s
Actual Super Pi time at 3194.75 mhz: 28.297s
Scaling efficiency: ~109%

I have no idea how or why my Brisbane scaled at over 100% efficiency, though it might have been due to an uncharacteristically poor Super Pi result at 1.8 ghz. However, I ran it without any major background apps (just Notepad and ClockGen) so it should not have been too far outside of the norm.

Nevertheless, we can conclude that the lowly Brisbane consistently scales well at high clock speeds unlike the E6300. Some of you might be crying foul about the memory speeds, but I would like to point out that the source article I linked above turned in nearly identical 1M Super Pi times for the 3.5 ghz E4300 w/ DDR2-900, the 2.66 ghz E6700 w/DDR2-800, and the 3.5 ghz E6300 w/DDR2-1000. Even with an extra 114 mhz of FSB speed and an extra effective 100 mhz memory speed, the E6300 fails to markedly outperform the E4300. This tells me that the Core 2 chips, at that speed, are limited by their memory controller more than anything else (in Super Pi anyway). The extra L2 cache of the E6700 obviously mitigates that limitation.

K8s like the Brisbane, however, love fast memory! Conventional wisdom has always dictated that you don't need fast memory thanks to the integrated memory controller. However, the integrated memory controller is the precise reason why K8 can successfully utilize fast memory. It is constantly bandwidth-saturated, so bandwidth is not the point . . . the issue at hand is latency, at least for the K8. By using fast memory running far above the HTT speed, you can guarantee successful 100% scaling efficiency at any clock speed.

I tested my Brisbane at the stock 1900 mhz (200x9.5) with the 1:2 memory ratio and received a memory clock of DDR2-760 5-5-5-15 2T. At that speed the memory latency (in cycles) was ~112 cycles according to CPU-Z. At 2761 mhz (290.63x9.5) with the 1:2 memory ratio with a memory clock of DDR2-1104.38 5-5-5-5 2T, the memory latency was . . . 112 cycles. It never changed. So long as the memory ratio remains the same, the processor scales upwards at 100% (or better) efficiency.

I sought to establish this point because I estimated that, at clock speeds above 3 ghz, the cheaper Core 2 chips (Allendales and 2mb l2 Conroes) lose a lot of ground to the K8 architecture due to scaling. In apps not so heavily slanted in Core 2's favor, K8 might actually be only 5-10% slower at the same clock speed. I was hoping to get my Brisbane up to 3.5 ghz to do some head-to-head testing with an E4300, E6300, or E6400 at 3.5 ghz, but the chip is simply not cooperating.

If I can get it to play ball, I'll try running some other benchmarks at 3.2 ghz later and see what happens. However, it looks to me like AMD could have competed with Core 2 had they actually released K8s in the 3.2-3.6 ghz range. It is unfortunate that they do not seem able to do so . . . even overclocking a K8 to those speeds is a daunting task.

 

[DrMrLordX]

Here are some additional benchmark numbers for the chip. Note that, in order to complete some of these benchmarks, I had to back off the 3.2 ghz mark (the vcores were scaring me a bit):

Testing done at: 3100.65 mhz (9.5x326.38) DDR2-1033.56 5-5-5-15 2T

CrystalMark ALU: 22678
CrystalMark FPU: 24702
CrystalMark MEM: 18031

PCMark05: 4550

ScienceMark:
Moleculare Dynamics: 52.753s
Primordia: 212.353s
SGEMM Method C: Peak MFLOPS 9057.98, FLOPS/cycle 2.92
DGEMM Method C: Peak MFLOPS 4466.68, FLOPS/cycle 1.44
Cipher Bench AES: 8.915s

CineBench9.5:
Single CPU: 49s 452 CB-CPU
Multi CPU: 27s 826 CB-CPU

 

[jaykishankrk]

Originally posted by: DrMrLordX
Here are some additional benchmark numbers for the chip. Note that, in order to complete some of these benchmarks, I had to back off the 3.2 ghz mark (the vcores were scaring me a bit):

Testing done at: 3100.65 mhz (9.5x326.38) DDR2-1033.56 5-5-5-15 2T

CrystalMark ALU: 22678
CrystalMark FPU: 24702
CrystalMark MEM: 18031

PCMark05: 4550

ScienceMark:
Moleculare Dynamics: 52.753s
Primordia: 212.353s
SGEMM Method C: Peak MFLOPS 9057.98, FLOPS/cycle 2.92
DGEMM Method C: Peak MFLOPS 4466.68, FLOPS/cycle 1.44
Cipher Bench AES: 8.915s

CineBench9.5:
Single CPU: 49s 452 CB-CPU
Multi CPU: 27s 826 CB-CPU

How can u r memory go so high :Q, is it a micron D9 chip based??

is u r OC stable for all the reading that you have shown??

anyway nice post though, keep it up

 

[myocardia]

Hey man, your numbers are definitely too low. My Opteron 170, at only 2.70 Ghz is faster in some of those benches, and not much slower in others. For instance:
ScienceMark Molecular Dyn- 50.249 seconds
ScienceMark Primordia- 242.896 seconds
ScienceMark DGEMM- 3923 Peak MFLOPS, 1.45 FLOPS/cycle
ScienceMark Cipher Bench AES- 10.468 seconds

Now, Cipher Bench is the only one where you're ahead by roughly the amount you should be ahead. I'm thinking that you aren't giving your cpu enough vcore. I've noticed with other benchmarks (although I can't remember if ScienceMark was one of them) that not giving your cpu enough vcore will actually give you lower scores, even though your cpu is running faster. See what happens when you leave your vcore the same, and lower your overclock to ~3.0 Ghz. You might be surprised.

 

[DrMrLordX]

Originally posted by: jaykishankrk

How can u r memory go so high :Q, is it a micron D9 chip based??

It's probably D9-based, but I'm not 100% sure. The base rating is DDR2-800 4-4-4-12 1T at 2.2v, and OCZ still covers it under warranty at up to 2.4v. The absolute highest I've push the memory is around DDR2-1130 5-5-5-15 2T which requires 2.4v.

is u r OC stable for all the reading that you have shown??

I've gotten the chip stable in everything at up to 3.15 ghz. Orthos was not happy at 3.2, but I have completed dual Super Pi runs at that speed.

3dMark06 seems to hate on the 6150 and produces all kinds of weird results. I'm not exactly surprised by that. Gonna try 3dmark05 later to see if it is happier.

anyway nice post though, keep it up

Thanks.

Originally posted by: myocardia
Hey man, your numbers are definitely too low. My Opteron 170, at only 2.70 Ghz is faster in some of those benches, and not much slower in others. For instance:
ScienceMark Molecular Dyn- 50.249 seconds
ScienceMark Primordia- 242.896 seconds
ScienceMark DGEMM- 3923 Peak MFLOPS, 1.45 FLOPS/cycle
ScienceMark Cipher Bench AES- 10.468 seconds

Now, Cipher Bench is the only one where you're ahead by roughly the amount you should be ahead. I'm thinking that you aren't giving your cpu enough vcore. I've noticed with other benchmarks (although I can't remember if ScienceMark was one of them) that not giving your cpu enough vcore will actually give you lower scores, even though your cpu is running faster. See what happens when you leave your vcore the same, and lower your overclock to ~3.0 Ghz. You might be surprised.

I'll give it a shot and see what I can do. However, keep in mind that my chip has half the l2 of yours and the l2 cache latency is higher. AMD claims Brisbane has 14 cycle l2, but both CPU-Z and Sciencemark claim it's 20-cycle.

btw, what is the memory latency (in cycles) on your Opteron?

 

[DrMrLordX]

I ran Sciencemark at 1.5v, 1.475v, and 1.45v to see if there were any major differences. I think my original run was at 1.525v or something unnecessarily high like that:

ScienceMark Trial 2 (1.5v):
Molecular Dynamics: 51.685s
Primordia: 209.317s
SGEMM Method C: Peak MFLOPS 9021.62 FLOPS/cycle 2.91
DGEMM Method C: Peak MFLOPS 4454.82 FLOPS/cycle 1.44
Cipher Bench AES: 8.955s

Sciencemark Trial 3 (1.475v):
Molecular Dynamics: 43.193s
Primordia: 210.208s
SGEMM Method C: Peak MFLOPS 9106.82 FLOPS/cycle 2.92
DGEMM Method C: Peak MFLOPS 4483.08 FLOPS/cycle 1.44
Cipher Bench AES: 8.933s

Sciencemark Trial 4 (1.45v):
Molecular Dynamics: 43.424s
Primordia: 209.803s
SGEMM Method C: Peak MFLOPS 9082.60 FLOPS/cycle 2.93(??)
DGEMM Method C: Peak MFLOPS 4472.38 FLOPS/cycle 1.44
Cipher Bench AES: 8.935s

Note: the Molecular Dynamics test seems to have a bug/glitch in it that will occasionally cause it to record completion times 9-10s higher than normal. I observed the clock jumping from 0s to around 10-11s in the space of one second, which was . . . odd. I think my original Sciencemark bench run had one of these results which is why my processor may have seemed to be so slow.

 

[myocardia]

Originally posted by: DrMrLordX

Originally posted by: myocardia
Hey man, your numbers are definitely too low. My Opteron 170, at only 2.70 Ghz is faster in some of those benches, and not much slower in others. For instance:
ScienceMark Molecular Dyn- 50.249 seconds
ScienceMark Primordia- 242.896 seconds
ScienceMark DGEMM- 3923 Peak MFLOPS, 1.45 FLOPS/cycle
ScienceMark Cipher Bench AES- 10.468 seconds

Now, Cipher Bench is the only one where you're ahead by roughly the amount you should be ahead. I'm thinking that you aren't giving your cpu enough vcore. I've noticed with other benchmarks (although I can't remember if ScienceMark was one of them) that not giving your cpu enough vcore will actually give you lower scores, even though your cpu is running faster. See what happens when you leave your vcore the same, and lower your overclock to ~3.0 Ghz. You might be surprised.

I'll give it a shot and see what I can do. However, keep in mind that my chip has half the l2 of yours and the l2 cache latency is higher. AMD claims Brisbane has 14 cycle l2, but both CPU-Z and Sciencemark claim it's 20-cycle.

btw, what is the memory latency (in cycles) on your Opteron?

I just noticed that you had replied to this.:Q I had to actually search for it through Google, so I could link to it, because of the lapping video. Anyway, my Opteron isn't a Brisbane, it's a Skt. 939 Denmark. That's why I was surprised that you weren't seriously outperforming me, especially at a considerably higher clockrate. Did you ever get it all figured out?

edit: Or do you think that it's just because of the higher latencies that are involved?

 

[DrMrLordX]

Hah looks like it took me awhile to notice your reply here too. Fear my thread resurrection.

The main thing I noticed is that an apparent bug in Sciencemark (hey, it is a beta) causes peculiarly high molecular dynamics scores occasionally. The "real" score of 43s makes my chip look much faster in comparison than the 53-54s scores I was turning in earlier.

Otherwise, I'm guessing the fact that I have half the L2 and slower L2 to boot is making your Opteron look better in comparison than you might think.

And yeah, I know you had a Denmark. I was just stating some of the flaws of the Brisbane core to possibly explain why my chip might be more sluggish than expected. Still, I think it did pretty well considering the price-point. Scales well, too. Had AMD rushed these out with high-k dielectrics and metal gate transistors, you might see them hitting speeds of 3.6 ghz and kicking all kinds of butt. Hah. Wishful thinking.