Does anyone know much about the Hypertransport Multiplier and how it affects things in this type of rig?

I read that it should never be set over 1000mhz.

I'm assuming that I should drop the HT multiplier from 5x to 4x when I overclock the FSB to 220mhz, because it automatically drops to 3x (I think, because Everest tells me the HT speed is 660mhz in that scenario).

If I manually set the HT to a 4x multiplier, the setting sticks, is that the correct way to configure it?

Any extra info is greatly appreciated.

Thanks Paul,

I did register and looked thru some threads, seems the Hypertransport link doesn't concern the overclocking all that much, but I've got to learn more.


Congo

Its actually quite simple.... The Hypertransport Mx should remain as close to 1000 (real) as you can get without going over, however, anything over 800 is smoking as long as you have tuned the rest of the system correctly. You should NOT try to go over 1000 if your CPU and memory are properly maxed out as this can blow a system apart, especially RAID0

Unlike the Athlon of the past, the A64 CPU has the northbridge incorporated into it. Therefore, using the system divider for the memory has absolutely no effect on performance as it would with non-A64 processors and it should be used if needed!!!!!!!!!!

The process for PROPERLY overclocking an A64 is different from the Athlon of the past. In a nutshell, here is the process

CPU:
turn off AMD C&Q
1. Drop CPU Hypertransport from 1000 to 200 (1x)
2. Set memory to DDR200 (reduce it to rock bottom for now.. you must remove the memory from the loop by running it well underclocked)
3. Set the vcore (3700+ san diego) MANUALLY to 1.5v (slight increase from the stock 1.4v to find the highest stable FSB)
4. Set multiplier to stock: (3700+ san diego) = 11x
5. Start increasing FSB. This will require you enter the BIOS after each Windows test using a CPU stress test software. There are many available. What you are looking for is 100% stability, with the highest FSB running the STOCK multipier (11x)

Lets say for this example you found that the highest stable CPU speed to be 2700 (2.7g) @ 11x

Next, the memory must be tested for max FSB. In order to do that, we want the CPU out of the loop like we did with the memory above.

With A64... ALWAYS set the memory to 1T CMD. YOU WILL SEE MORE 3D PERFORMANCE WITH 1T ENABLED THAN WITH ANY OTHER MEMORY TWEAK. Your better off running 1T @ 230fsb and tighter timings than running 2T @ 270fsb with loose timings!!!!!!

Mine is at 211fsb @ 1T running the 5:3 divider @ 250x11 CPU
Benchmark Results
RAM Bandwidth Int Buff'd iSSE2 : 6651 MB/s
RAM Bandwidth Float Buff'd iSSE2 : 6606 MB/s
(I am limited by AGP motherboard A8V Deluxe)

Memory:
1. Set the CPU multiplier 4x less than stock (3700+ san diego 11x - 4x = 7x)
2. Leave CPU Hypertransport at 200
3. Set the memory ratio to DDR400
4. Set the FSB to 200 to start (or start where you know the memory is stable.. ex:230, 235... etc)
5. Same as the CPU, increase FSB by 5, boot into windows and run memory stress software untill errrors start to appear. At that point you have a choice... lower fsb, change memory timing or increase Vdd. If your memory is rated for 2.6v, bumping Vdd up to 2.7 or 2.8 is acceptable but lower is better. The most important timing is the CAS rating and the TRAS however raising CAS 2.0 to 2.5 will not hurt performance... CAS 3.0, will. Rasing TRAS 2-3-4clks is also OK.

Once you have found the highest stable FSB with 1T-CMD enabled, your ready to calculate the overclock. Lest say for this example you found the max FSB for the memory to be 250fsb... So the MAX stable CPU speed is 2700 and the MAX stable memory speed is 250

Now that you found the max FSB of your ram and the max speed of your CPU it should be a little easier to clock it together..

Lets say your CPU maxed out at 2700mhz with its default multi of 11x.. But your ram maxes out at 250mhz..

11x 250 = 2750mhz.. That won’t work (without raising Vcore)

10x 250 = 2500mhz .. That will work but your cheating yourself 200mhz of CPU power.

11x 245 = 2695MHz .. That will work and get you closer to the max of your CPU. And running the ram little slower then the max you found before to achieve the higher CPU MHz might also allow you to tighten up the ram timings a little.

Now there is always the option to use a divider for your ram which will get you closer to your CPU and RAM’s max....

10(CPU multi) x 270(FSB) = 2700mhz (MAX tested CPU speed)

and if you use a 9:10 divider with this config…

270/10  x 9 = 243mhz (mem speed) ... and you can reduce memory timing or vdd because the overhead of 7mhz is there!

With whatever you decide, there really isn’t a right or wrong way to do it. The only way to find out which way is best for you to run is test but do not discount the divider with A64 systems!!!

LAST...increase the Hypertransport number until it is closest to 1000 in Windows without going over 1000. In some cases this can be set to 600mhz in the BIOS and still be well over 900mhz+ because of the overclock you established above.

Mine is at 800mhz in the BIOS which equates to exactly 1000mhz with my overclock:

Front Side Bus Speed : 2x 1000MHz (2000MHz data rate)
In/Out Width : 16-bit / 16-bit
Maximum Bus Bandwidth : 8000MB/s (estimated)

That is how Hypertransport works and how it relates to the system in overclocking. It is the northbridge communication rate and anything over 800mhz is VERY ACCEPTABLE when overclocking.... Truth is, anything over 500mhz you will never see ANY loss of performance in your system.

Hope that helps.

PS:
Vcore is the key to stable A64 overclocking... this can go as high as 1.65v (without extreame cooling) with a 3700+ san diego although I dont run mine maxed out, (mine is at 1.5v and 250fsb). It must be set MANUALLY as well as setting the CPU multiplier MANUALLY to overclock correctly. Once those are set manually in the BIOS you can re-enable AMD C&Q (cool & quiet). With C&Q enabled the settings are usually not available in the BIOS but are still ACTIVE as the base settings for C&Q to use.

Quote:



Your welcome!

With the older Athlon/Duron/Sempron the system divider can have a negative effect on performance. With the A64, the divider has no effect on performance and will allow the user to significantly enhance the system.

It was easier in terms of testing time to overclock the older processors mainly because the overclock was relative to all components. With the A64 it requires overclocking each system (CPU/Memory) independantly, testing settings for hours, then calculating the final overclock values and testing again. The results far exceed the Athlon of the past since it is not unusual to gain 600mhz+ on a A64 CPU, ESPECIALLY the San Diego core. If you max out the vcore and use extreame cooling... an OC of 800mhz+ is very possible.

A64 systems respond to the 1T CMD memory setting with close to TWICE the performance gain of previous processors. That is why it is best to either select memory rated for 1T-CMD or set the 1T-CMD and work out the memory timing/FSB/vdd based on that being enabled, even if the memory is not 1T rated.

I do alot of video editing and use 2 1G sticks of OCZ 2.0-3-2-5: 2T memory rated at 2.5v. By raising CAS to 2.5 and running 3-3-8 (or 11) with a vdd of 2.7v and the divider of 5:3 (DDR333) I can run the 1T-CMD stable up to 260FSB CPU with a Vcore of 1.65v (216mhz memory), which nets over 7000 in memory scores. When I push the system like that I drop hypertransport in the BIOS to 600mhz, which nets an HTT overclocked rating of around 845mhz.

With that overclock I am well over $1000 dollar FX57 performance using a $245.00 CPU.

The reduced 845mhz OC/HTT will not have any negative influence on performance, at all.

1000 is the target value you do not want to go over but as long as it is over 700-800 and the rest of the system is correctly overclocked... your system is smokin!!!!!!!!

I've never tried Cool & Quiet or anything automatic, but I was going to try it just to test it out.

A lot of great info there Nick, I learned more from those few posts than I have looking around the overclock forums for years, I never seem to read  a post that's relevent to what I wanted to know, it seems they talk about every single aspect except how to actually do it. Even the guides I've read were hopeless.

Another thing that happens a lot, when I use methods proscribed, I don't usually get the results I'm expecting, this really throws me off and creates even more confusion. This happens particularly with my ram settings and tests.


I got over 7000mbs/sec pretty much straight away on my overclock, and then everyone tells me I'm doing it wrong, so wtf is up there? Isn't the whole point to increase bandwidth?

Is there any way to figure out how to set all the complex ram timings, not just the fabulous four?

My Asus A8N SLI board does some crazy stuff with my settings most of the time. I have it all set to manual, yet the darn thing keeps modifying my settings, in particular my HT multiplier, I find it difficult to get that setting to "stick" even when I know that it is correct.

My memory divider settings are not that great as I found out. I get  the divider by selecting 433, 400, 333, 266 etc as the base speed. I wish there were some more intermediate ratios, I believe some DFI boards have them. My mobo will auto set the divider and HT multiplier 2 steps down if it doesn't like a setting, and this sucks. It also modifies my ram timings when it feels like it. I had trouble with these things when I first got this board, but have been working around them, mostly by luck and shear experimenting. I'm having problems mapping consistent behaviour though. This has been the same, with slight variations, through 4 or 5 bios updates, which seem to be coming out regularly. I'm guessing they haven't yet sorted the bios for these boards, and it may take a while.

The most important starting point is the BIOS. Finding out which BIOS is most stable for the purpose of overclocking, which ones are known for issues and what the issues are. Asus is well known for releasing a motherboard before it has been fully tested and fixing the problems in BIOS releases. Reviewing the forums for BIOS upgrade/downgrade information is the one of the most critical steps allot of people skip and as you have found out it can also be a headache. One of the reasons I selected the A8V Deluxe was because of the BIOS/stability research I did prior to the purchase. I considered upgrading to a PCI Express video card and SLI system but after reviewing several motherboard selections I found that the expense may not provide the results I was after, so since I already had an AGP-ATI X800 AIW I elected to stay with an AGP board for that purchase. I will probably make the PCI-E plunge after the next round of cards/motherboards have proven themselves for at least 9-12 months on the market. I very rarely purchase a motherboard or card that has been on the market for less than a year as I do not wish to be the research and development department for thier products.

Not to post a long winded techno-head answer, but in response to your questions the bottom line is:

1. If you intend to "truly" overclock, in every sense of the word, liquid cooling is a must for ALL CPU/GPU components, period. Heatspreaders are also required for the memory.

2. If you intend to performance tune, there will be limits to how far you can go… and how far you need to go.

3. The memory purchase is the key to performance. Kingston memory is crap because of their inconsistent quality even with their top of the line products. Crucial and OCZ are the kings of consistent quality and support. How high memory can clock past its ratings is really hit-and-miss even with the better companies and it all comes down to the PCB layer design and the chips they use for the production run.

4. The clock gen IC's on the motherboard can get hot and epoxy a tiny heatsink on them can net 10+Mhz is memory and/or CPU FSB, or make a flaky motherboard rock stable. I will typically find my best O/C settings and run 3D stress tests while probing all the IC's on the motherboard with a digital thermometer to find the excessively hot IC's. Once located I will cool them with an attached heatsink. Same with the video card, a dab of epoxy and a chunk of metal has a significant affect on a system.

Every BIOS memory timing setting does not need to be tweaked. The CMD+FAB-4 (2.0-2-3-5) is all you need to deal with. The CMD (1) rating of the memory is the most important, followed by the CAS rating and T-RAS. A CMD of 1 with a CAS rating of 2.5 is excellent, CAS 2.0 all the better however it is entirely possible to run a 1-CMD on 2-CMD memory by manipulating the FAB-4 timings and Vdd.

Memory bandwidth scores are important although not the holy-grail. What is important is that the maximum stable CPU speed is obtained FIRST followed by the maximum stable memory FSB with 1-CMD enabled and remain at CAS 2.0 or 2.5. There is no such thing as a memory timing tweak that will provide any ‘significant’ increase in performance other than 1-CMD and the FAB-4. Leave the rest on “AUTO”

Vcore, multiplier, FAB-4 memory timing, Vdd, AGP/PCI (if applicable) settings in the BIOS must be set manually. Any “Auto” or “AI Overclocking” must be disabled.

The divider system is BIOS/Motherboard dependant and can force other areas of the system to be changed when enabled. This is something that is differs for each motherboard and/or BIOS and should be researched prior to purchase to ensure you are getting what you want in the way of control over the system.

When posting at overclocking sites to compare your system for analysis, always remember the following:

1. The person on the other end may be lying about their scores and specs (very common)

2. The person on the other end is a serious overclocking pro and they don’t always give away their secrets… remember it is a very competitive bragging rights hobby, just like auto racing.

3. The person on the other end happened to get a set of components that far exceeds the specs of those components for quality. I have personally built 3 duplicate systems and had all three perform with significantly different performance results using the same tower/PSU and BIOS's.

7000+ in memory benchmarks sounds great to me but without being able to do my own tests I do not know if that is good or bad for your system (with the components you are using). I would not put too much empahsis on what someone tells you about how poor that score is from a speed-freak website. Check over at PC Perspective and see if your overclock/benchmarks tend to fall in line with the systems being discussed that match your components.

Although old, I also use AquaMark 3 for a baseline performance anaylsis. That software is greatly influenced by memory bandwidth and provides a visual means of judging performance changes.

These may also help:


http://www.overclock.net/faqs/31782-how-do-i-find-limits-my.html

http://www.overclock.net/faqs/38169-asus-a8n-sli-motherboard-introduction.html

Quote:


To wrap it up…

With my A64 system, I can run CPU @ 9.5 x 295fsb (2802.5mhz) with a divider of 4:3 (DDR266), meaning my memory is running at about 187mhz with 1CMD enabled and HT set to 600Mhz. My memory NUMERICAL benchmarks are significantly lower (by 1000pts) and hypertransport sits at 2x885 or 1770 total.

With that configuration I must run the Vcore @ 1.7v and Vdd @ 2.8v, which is not a problem but I prefer to run a bit lower if possible. The bottom line is:

What are the real world performance gains?

With the above configuration my 3D benchmarks are HIGHER than if I run @ 11x250 (2750mhz) with a divider of 5:3 (DDR333), meaning my memory is running at about 211mhz with 1CMD enabled and HT set to 800Mhz. My memory NUMERICAL benchmarks are significantly higher (by 1000pts) and hypertransport sits at 2x1000 or 2000 total… and I get a LOWER 3D benchmark score.

What that means is the memory benchmark NUMBERS are NOT important. What IS important is maxing out the CPU with the highest FSB possible and finding the divider that allows it to boot and run with 1CMD enabled and the lowest memory timing possible.

Now… looking at it from another perspective, the difference in the performance gain is there, but the question is, …is it worth the gain??

In the above case running slower with the divider 187fsb (DDR266) increases Aquamark3 or 3Dmark tests by about 500-700pts, therefore the gain I got from running 9.5 x 295 @ DDR266 required voltage increases that in MY opinion are not worth the 500-700pt performance gain.

Unless I see an increase of 1000pts or greater in 3D application tests I (personally) cannot justify the stress of the voltage increase. If I was able to run the above without increasing Vcore from 1.5 (or even 1.6) to 1.7, I could justify the overclock. That is my own guideline as most overclockers would go for the higher 3D benchmark.

My point is, the numerical memory benchmark is not as important as the real world performance. High memory benchmark numbers create a placebo performance effect.

Is that always true?

No.

Motherboards/memory products/BIOS’s can be different and produce different results however more often than not I have found that high memory numerical benchmarks do not produce the best real world performance. The final test in overclocking or performance tuning is in the real word use of the system.

I will guarantee that an overclocker changes his/her BIOS to post the memory braging rights number, then switch the BIOS around to run other tests to get the best 3D test numbers. Its a game and unfortunately without a degree in EE or a comprehensive education in digital electronics, most are clueless about how that game is played online by the overclocker ego.

Quote:


ahhhhhhhh.... Yep, you should be using a PSU with a 12v rated at 28A minimum. The higher the overhead current rating (over what your system needs) for the PSU, the cooler a CPU will run and a higher stable overclock is much more possible to achieve.

As for 1T.... As an electrical engineer I disagree with that site in that conclusion. Although I understand the logic being applied, the real world performance is not the same as the benchmark and 1TCMD will always provide better performance IF the memory can run it. I would say from your posted timings/settings and being the memory you are using is PC4000 rated, that PC4000 rating comes from 2T, loose (high) timing, not from a true high performance curve... and in that you are correct that if you were to purchase TRUE PC4000 memory designed for overclocking, 1TCMD would smoke the memory you are using at lower FSB with tighter memory timing.

As I posted earlier... the memory purchase is the key to the entire overclock and system stability

What you achieved was the best with the quality of memory purchased and that may be why someone told you your 7000 memory benchmarks were considered low, which could be true because you cannot run 1T-2.5 or 1T-2.0 to test. If my 2x1gig PC32000 sticks can hit close to 7000 I would think your system would provide a 7500-8000 score or better with the right memory product, but I am not experienced with your motherboard so what I am saying here is based strictly on education and not specifics.

Quote:



What you are describing sounds very much like over-voltage damage. Progressive reduction of performance usually indicates damaged components from heat and/or voltage spikes. It may be the memory itself and not the memory controller or CPU.

PC3200 running @ true DDR600 (300FSB) sounds a bit far fetched to me. Assuming the motherboard would be stable for that speed, the memory would need to be run at very high timing with high voltage and very high efficency heat transfer in place.

Quote:



I guess I'd better clear up what I tried to say about the 7000mbs/sec memory bandwidth I achieved.

What I meant was, here I am testing after a few minutes with my new stuff, and I reached that amazingly high memory score with my inexperience, using the default 2T Command Rate and loose ram timings. I wasn't complaining about the score, I was questioning why someone commented on how I should go about it when I probably have a considerably faster system than they do, despite their professed superior settings.

Now, what I really wanted, was for them to justify to me, in evidence or anecdote, the reason why I should change to 1T and dividers and blah blah when I clearly provided graphic descriptions of my success complete with screenshots showing my various settings and achievements.

I want to know what's wrong with doing it my way if it provides results.

As it turned out, I think I've found the answers since then. For one thing, there is a different concept applied with the type of RAM I bought, it runs at lower voltages on the overclock, loose timings and a very high speed as standard.

I did manage to run and test the RAM at CAS 2 and 1T (as I flew dogfights online for 4 hrs tonight straight, the PC locked up as I left the game!) This was at 2.9Vram 2-3-3-6 1T , memory set to 5:6 divider, memory 220mhz (DDR440), 2886mhz on the CPU and HTT at 272mhz x 10.5

I managed to get a tiny improvement in my 3D MArk '05 score, a matter of a few dozen points. I'll retest using the other settings soon.

What does seem apparent, is that the CPU speed has a lot more influence than the memory bandwidth on performance in these new systems. I theorise that system bandwidth is more than adequate on this type of rig, so the only thing slowing it down is the CPU. This was reflected tonight in my 3D Mark score, slightly up and in proportion to the slightly higher overclock I managed to get stable on my CPU with lower memory speeds. I doubt very much that the 1T command rate and tighter timings at a lower memory speed had anything to do with it.

Well yes but at least on the NF4 where I am very familiar. I have yet to see a system on tight timings low fsp outperform high fspand loose tmings(on NF4) can't speak for any other systems. I cant belive the xbox360 is water cooled, wow I wonder how they pulled that off.