Can most nforce3 motherboards run the hypertransport bus above spec?

[PingSpike]

Where would instability from the hypertransport bus manifest itself? Memory? Hard disk? Does increasing the chipset voltage increase the stability in regards to this element of the motherboard?

 

[Edzard]

I'm running the Rev 1.0 version of the VNF3-250 right now with HT at 1000 (250 x 4). At least one on-line review mentions that the board does HT 1000 reasonably easily.

I have also run it at HT 960 (240 x 4) and HT 900 (300 x 3). The seond of these two settings was done to obtain 2.1GHz from my Sempron 2500+ processor (stock 1.4GHz), with memory set at 133MHz and Vmemory set to 2.7V (Nanya PC3200 memory w/ stock voltage of 2.6V)

I have installed a 45 x 45 x 10MM fan on the NF3 chip and suggest you install a fan on your NF3 chip, ESPECIALLY if you intend to increase the chipset voltage.

I'm thinking that instability from the HT would be manifest in memory. If I'm right, that instability in memory could easily ripple through the system to impact your hard drive.

 

[PingSpike]

Thanks for the input Edzard. I believe I have revision 3.0 of this board. I'll try bumping it up this weekend, and will try running the prime memory stress and memtest86 on it at that point. I've been a bit nervous about overclocking this board on the stock heatsink since the temp sensors seem useless. It tops out at 50c, even at stock...ramping up the speed nets the same temps. I've heard from others the temp sensor on this board is pretty useless...and I'm still at stock vcore so I haven't really worried about it.

How did you go about mounting the fan on the chipset? Due to laziness at this point, I'd rather not remove the heatsink if at all possible.

 

[Edzard]

I wonder if the temp sensor is "pretty useless" or if the problem lies with the software that reads it. With the V6.0 BIOS, temperature readings displayed in the BIOS setup and by Chaintech's "DigiDoc" monitoring software are more plausible than those reported by earlier BIOSes but don't have perfect agreement by any means. At elevated temperatures resulting from OCing, on my Rev. 1.0 board, "DigiDoc" seems to report temps that are lower than those reflected in the BIOS setup, and seem to max at 49 deg. C. I'm using a stock HS/fan in combination with Arctic Silver's Ceramique thermal paste.

To date I've been running this board in a half case obtained by what the Brits call "bin diving" at a local corporate PC recycler's place. It has no top, bottom, and lacks one side - a great test stand. While running in overclocked mode, I have frequently touched the side of the CPU's heatsink and never detected any temperature that even approached "hot". However I must note that the design of the heatsink mounting bracket on the board is such that I am unable to touch/sense the temperature at the lowermost base of the heatsink (which is of course closest to the processor).

Due to the results of some experiments I conducted yesterday, I am now convinced that I have NOT run my board "at HT 960 (240 x 4)" as earlier posted, even though that was what I had input and in fact was shown in the BIOS setup. I'm sure that the BIOS actually dropped that setting to 240 x 3 without giving any indication of having done that. On the other hand, I'm equally convinced that the BIOS did allow me to run at 300 x 3 (memory set at 133MHz) to obtain the 2.1GHz OC. In the near future I'll be posting the results of those experments which I think VN3-250 enthusiasts will find quite interesting.

With regard to installation of the fan on the motherboard's NF3 chip, I simply mounted a fan on the stock heatsink, without removing the heatsink. The fan I used dimension-wise is almost exactly the size of the footprint of the top of the heatsink. Attached the fan with two small tapered "woodscrew" type screws, screwed into the gaps between the heatsink's vertical cooling "fingers". The design of the heatsink's attachment to the motherboard negated the use of four screws (one at each corner of the HS). But I have frequently tried to wiggle the fan to see if it is vibrating loose - it hasn't so far (installed about 5 days ago) and I don't think its going to. I advise that if you use a fan that is a near-perfect fit to the top of the heatsink (as I did), then choose the two mounting screws carefully to ensure that they will "bite" into the cooling fingers well to produce a secure/solid mounting.

That's all I have time to say right now since I am determined to go "live" with my VNF3-250/Sempy 2500+ setup this weekend. It will replace an ABIT BX6 R2.0 with a Celeron 600 OC'd to 900 that served me well for five years and in fact is still running.

Good luck with your VN3-250.

 

[PingSpike]

Yeah, I'm unsure about the digidoc software...I think I had a similar experience with regards to the temps in the software versus the bios but I can't remember for sure.

What makes you think the HT wasn't running at the settings you specified?

I also did the "touch test" on the CPU and had the same sort of opinion you did. It definately wasn't hot, but with nothing to really compare it to I couldn't say it was good either. I can't exactly remember being burned when touching a heatsink before. 😛

I'll see about the northbridge fan.

 

[Edzard]

Originally posted by: PingSpike

What makes you think the HT wasn't running at the settings you specified?

By making two sets of runs of the SiSoft Sandra 2005 benchmark (benched CPU, multimedia, and memory), with BIOS settings of CPU @ 250MHzk and memory @ 166MHz for both runs, with HT set at 3x for the first run and at 4x for the second run. The results were, for all practical purposes, identical (differed by a maximum of 0.14 percent [CPU Whetstones]) - within the variance one would expect to observe when making two consecutive runs of Sandra keeping all settings constant for both runs.

 

[PingSpike]

Originally posted by: Edzard

Originally posted by: PingSpike

What makes you think the HT wasn't running at the settings you specified?

By making two sets of runs of the SiSoft Sandra 2005 benchmark (benched CPU, multimedia, and memory), with BIOS settings of CPU @ 250MHzk and memory @ 166MHz for both runs, with HT set at 3x for the first run and at 4x for the second run. The results were, for all practical purposes, identical (differed by a maximum of 0.14 percent [CPU Whetstones]) - within the variance one would expect to observe when making two consecutive runs of Sandra keeping all settings constant for both runs.

Ah. Well that might be in line with the norm actually. This article seems to show that hypertransport bus speed has a pretty negligble effect on performance, at least until you get down to 200mhz. (and even then, its not much) In the case of the game benchmarks at the end of the page, 400 and 600mhz hypertransport settings were outscoring 1000mhz, suggesting that in those particular applications its effect is so small its within the margin of error. (to me at least)

I heard that the hypertransport bus had little effect on performance a lot, but I usually like to see some kind of justification for these statements for myself. When I dug around I came up with that article. Its pretty good, but I would have liked to have seen them run a few more game benchmarks, some with higher settings. I understand why they did what they did though.

 

[Edzard]

Thanks for pointing that article out. Looks like I may have to eat crow re my very first post AND my last post. The effect of varying HT is so small that my allegation about Chaintech's vnf3-250 and 3x/4x either isn't valid or not provable, at least with the bench I used.

However, the BIOS does appear to be examining the settings. I say this because my rig essentially won't boot at 300/133/2x<== (CPU defaults to stock 1.4GHz frequency and a message appears instructing me to enter the BIOS setup and check settings). But my system WILL boot without complaint (at the expected 2100MHz = 7 * 300) with settings of 300/133/3x<===.

Since you are also running a Chaintech VNF3-250, please try this out: set CPU at 250MHz, memory at 166MHz, and HT at 3, run the CPU-Z utility and look at the *memory* frequency. Then set CPU at 257MHz and do not change the memory or HT settings. Again, look at the memory frequency as reported by CPU-Z. Then please post details about what processor you are running, what its core revision is (shown in CPU-Z), and what your memory frequencies are at the two settings.

On my board, CPU-Z reports memory at 194.4MHz at the 250MHz setting (I expected 208MHz) and 200.2 at the second, 257MHz CPU setting. So Im running my rig at 257/166/3 to obtain 1800 CPU and 200.2MHz memory frequencies. This is what I was mainly alluding to when I stated, in an earlier post, "In the near future I'll be posting the results of those experments which I think VN3-250 enthusiasts will find quite interesting." But perhaps this apparent memory underclocking associated with CPU overclocking is already known, perhaps CPU-Z isn't corectly reporting the memory frequency, or something else. I've had to eat a lot of "humble pie" lately (pie, crow - take your choice) .....

 

[PingSpike]

The A64 is a weird platform to learn compared to its predecessor, I've just taken my first steps into it in the past few months. To further complicate matters, there's a mess of misinformation and no standard set of terms for anything. (CPU clock is refered to as tons of different things) It took me awhile to figure out how how memory dividers seem to work, every guide I read seemed to confuse me even more. Those results seem in line with what I would expect them to be. My machine is at home so I can't confirm for sure.

The clue I used to make my decisions was in cpu-z, where it will show the divider used. It shows up as cpu/x. The divider I've heard has to be a whole number because of the way A64 works, so no 11.5 or 12.5 are used. It can't be manipulated directly, and it chosen as a function of *default* CPU speed. This divider can be manipulated indirectly by choosing a desired memory speed in the bios as you know.

So for yours, 1400mhz is the default speed.

200 memory frequency (400ddr) = 1400 / 200 = 7 = rounds to 7
166 memory frequency (333ddr) = 1400 / 166 = 8.43 = rounds to 9
133 memory frequency (266ddr) = 1400 / 133 = 10.526 = rounds to 11

It always seem to round up to the nearest whole number, I think they're trying to be conservative for stability. I think the settings are part of the cpu itself. I just sort of found this out on my own through testing, which is what one of the guides that confused me said I would have to do to understand it. It works for your example though. 250 * 7 = 1750mhz / 9 (166 setting) = 194mhz memory clock.

I like that A64 platform overclocking its versatile and rewarding, but its kind of a pain in the ass at first. Beats being stuck with a up and down locked multiplier and a fat performance hit from running memory async like in the late Athlon Xp days though.

Thats really odd that the bios wouldn't boot using 2x but would with 3x though...underclocking usually doesn't cause any problems! I have no idea what the deal there is.

 

[Edzard]

Thanks for the elucidation. I'm getting absolutely stuffed with crow-pie.....

In an effort to repay you for the education you've given me, and in an effort to address your original question at the top of the thread, I googled for "hypertransport", glanced through several of the search results, and chose:

Decent primers:
http://www.ukgamer.com/article.php4?id=143&page=1 (save the pictures, then look at them - much better)
http://www.devx.com/amd/Article/26985
http://www.devx.com/amd/Article/27012
http://www.interfacebus.com/Design_Connector_HyperTransport.html

THE authoritative source:
http://www.hypertransport.org/

But you may well have already investigated these sources.