Hardware/Memory FAQ (mostly) by CandyKid - Read it, it's awesome-o!

[CandyKid]

If anybody has cryptic/odd/hard to find hardware questions, please post and we'll just keep adding answers!

DDR-II Memory

Architecture:

The second generation DDR memory is not just a speed upgrade (quite the opposite), but an entire systematic change.

I've seen quite a few questions about adding PC5400 modules to a i875 chipset board and had to dissappoint the user by informing them that it just won't work. Apart from having ~60 extra pins, DDR-II memory modules are also designed to run with a lot less voltage.

While some uber overclocks with DDR-I require as much as 3.3 volts, there are now DDR-II modules that will hit PC8000 (DDR-II 1000) with only 2.2 volts. In fact, the DDR-II spec is well below DDR-I (JEDEC is 2.5v for DDR-I) usually only requiring 1.8 or 1.9 volts for a fairly lofty system.

However, we are easily fooled into thinking that our DDR-II modules are running much faster than DDR-I and this is incorrect.

DDR-I basically stops at PC4400 while DDR-II pretty much starts at PC4300 and goes upwards to PC8000, but we can expect to see even higher ratings comming with newer chipsets!

PC4400 or DDR550 runs an actual clock of 275 MHz with two bits of data sent per clock, one at the rise and one at the fall of the signal (hence 'double data rate' memory). This doubling is why you see a 275 MHz rated module listed as DDR550!

DDR-II, however, has an integral change in the way the memory works. Instead of two bits per clock, it actually sends four. This means that the PC4300 DDR-II modules are actually running HALF as fast as the PC4400 DDR-I modules, but still maintaining the nearly same bandwidth!

The PC6400 (DDR-II 800) modules you can buy are actually running at 200 MHz, the same as PC3200.

In effect, DDR-II doubles the 'double' of DDR-I and allows us to get the same bandwidth at half the clock frequency!

Carry this thought onwards and the PC8000 modules we see don't look so daunting since it's only a 250MHz frequency (Only a 50 MHz OC for a stock 200 MHz system like the i955 chipsets!).

PLUR
CK

[CandyKid]

I havn't been keeping up with mobo technology. Are there any out there with ddrII support, or is ddrII near future technology

DDR-II support has been embraced by Intel and all their Socket T (LGA 775) systems support it (Chipsets: 915, 925, 945, & 955). Most of these boards natively support PCI-Express add-on cards and 16x video cards, but there are a couple of variations with an AGP slot or, rarely, both.

The only one to look out for is the i915 chipset as the lower-end versions don't have DDR-II support (above i915p does).

Expect to see nVidia DDR-II support with their nForce chipsets within the next two quarters as well as the ATi Express chipsets, the new BFG chipsets (They're making chipsets!?!?), SiS, and VIA, etc.

PLUR
CK

p.s. I don't follow SiS or VIA very much, do they already have DDR-II? SiS might... I don't remember without looking it up!

[CandyKid]

So, todays DDR1 is reaching speeds in excess of 300MHz in the right hands. DDR2 reaches 250MHz (is there overclocking potential above this?). Will the DDR2 modules be limiting the increase of the HTT bus / FSB for the CPU?
I havn't personally seen DDR-I modules running at 300 MHz (DDR600/PC4800) and the highest rated modules you can buy off the shelf are 275 MHz (DDR550/PC4400), but that doesn't mean it doesn't exist! I'll catch someone running a 2:3 divider now and again, but something usually becomes unstable when pushing the system. Take note, however, that the dividers, etc, seem to work a bit better with DDR-II.

DDR-II modules are still very young and 250 MHz (DDR-II 1000/PC8000) on the actual clock isn't pushing the registers to their maximum. I highly doubt the modules will be a limiting factor.

if the memory is actualy clocked slower than regular ddr why is there such a fuss over the ddr2 being hot i mean shouldent the lower clock speeds create less heat?

The only reason I can think of this, is that its drawing more current. Current is actually what causes more heat than voltage (iirc, and I havn't taken courses in applications of circuitry yet, so Im not sure, just going on general physics assumptions.

DDR-II memory is fundamentally different in the way that it works and functions. Yes, there should be a little more current, but it is also doing more work per clock cycle, producing more heat.

also just something i have read AMD has chosen to not support DDR2 yet so im guessing you wont see an Athlon 64 or X2 with ddr2 support anytime soon.

AMD has always been notable for their efforts in improving clock for clock performance rather than going for pure speed. Their CPU data paths are extremely short compared to Intel which is why a 2.2 GHz A64 will be on par with a 3.6 GHz Pentium. AMD has also made a lot of headway with their memory cache programming, prefetch commands, and logical accesses all of which make their AMD 64's very proficient at maximizing the memory bandwidth.

Unfortunately for them, they're going to max out with DDR-I because of their low clock speeds/MCH frequencies and the cieling of DDR-I. The other fact they have to face is that manufacturers are phasing out DDR-I IC's (memory chips) and thus they'll go the way of SDR SDRAM, eventually.

AMD doesn't really have a choice. What they're doing, IMHO, is perfecting the technology first and waiting until it's absolutely necessary because when they do switch over, the MCH's in their A64's are going to be completely revamped for DDR-II memory functions.

PLUR
CK

[CandyKid]

What's the big deal with PCI-Express?

PCI systems have 132 MB/s of bandwidth. Total. So if you've got a RAID card, an ethernet NIC, a sound-card, and a PCI USB port card (adds 4 ports), then you're using 132 MB/s for ALL OF THEM AT THE SAME TIME...

If you add in a PCI graphics card, you're dead out of bandwidth.

The other problem is that the old PCI architecture doesn't support a 'full-duplex' (remember that from NIC's) and can only send data in one direction at a time. This even applies to the AGP slots (which, incidientally, do give you a relationship to the PCI slots... AGP 1x is 2x faster via a 66MHz frequency).

A 1x PCIe slot delivers a maximum 250 MB/s bandwidth in EACH direction, meaning that the cards can talk back to the system (requesting information, etc) without interrupting their own data stream (I.E. output cards like video). This is ~ 2x faster in only one direction!

Add in to that mix more lanes and you end up, eventually, with the x16 slot for graphics, delivering 4GB/s bi-directional bandwidth making it twice as fast as an AGP 8x (~2.1 GB/s uni-directional). There is currently a specification (but no implementation) of a 32x PCIe slot and at 8 GB/s in each direction, we might need to do something about the memory subsystem so that it can keep up (Hell, DDR-II 800 would only just be there!).

I guess what I'm getting to in all this mostly useless info is that PCIe offers more bandwidth that is finally dedicated. This means we can go faster and do more with each slot and although hardware has been slow to catch up (as said, because they don't need it... yet), they will soon enough. For now, focus on your video as a mainboard can always be upgraded! :p

The savvy among you might have a thought tickling the back of your head right now... "I've seen this before."

Yes, you have. PCIe is a serial setup, unlike PCI which is parallel... do you remember RDRAM from Intel? Same thing. It was serial, but had the very unfortunate problem of having a really bad controller and poor implementation causing a serious loss of performance the more you added to your system.

PCIe has addressed nearly all of the issues at hand and is ready, in fact, for you to add cards into the empty slots. PCIe, however, has an advantage here because the mainboard manufacturer knows exactly how many slots you have and can thus plan for the additions with their architecture. The RDRAM team at Intel seemed to plan their system on very lean computers with only a video card and MAYBE one additional PCI card.

PLUR
CK

Reposted, but applies to this thread.

[CandyKid]

I just noticed this on the front page of tomshardware. It talks a little bit about ddrII, but focuses on pushing the limits of a current DDRI chip slated to handle extremely tight timings, or high clock speeds. Tom's Article (http://www.tomshardware.com/motherboard/20050721/index.html)

I wouldnt' use them on my system and I'd be willing to bet they don't cover damage done to your hardware due to high voltage, eh?

Besides that, the difference isn't that much. Between CAS 1.5 and 1 would be a nice jump, but between 1.5 and 2 isn't... same as between CAS 2 and 2.5 v.s. 2.5 and 3! Heheh... oh man, that'd be a whole different thread to explain that one!

Let it suffice that a half clock does almost as much damage as a whole additional clock.

PLUR
CK

[CandyKid]

Found a question: Which memory modules have Samsung TCCD IC's

This is a DDR-I related question, but an easy answer.

Modules listed as 2.6-2.8 volts which run the latencies of 2-2-2-5 @ DDR400/200MHz are usually Samsung TCCD IC's.

They used to be Winbond -5 Rev B IC's and if you ever see modules rated as 2-2-2-6 which are a first revision, that's probably BH-5's.

However, nowadays, there are also PC4400 modules and those listed as CAS 2.5 and 2.7-2.9 volts are also usually TCCD IC's proving their overclocking prowess and usefulness especially for the crowd that doesn't want to overvolt their RAM way out of spec.

Be careful, however. Some companies such as OCZ list memory modules with different ratings based on voltage. If they require 3.x+ volts for a 2-2-2- latency, then they are not TCCD IC's or there is something wrong with the module design.

UPDATE:

Samsung TCCD IC's are actually a custom specification made specifically for Corsair Memory, Inc. So while the IC's are becomming a little more rare around the world, Corsair has stated that they will have a garaunteed supply well into next year.

PLUR
CK

[CandyKid]

Something I've had to hit on a few times in the last few weeks:

Properly balancing your computer hardware.

Lately, I've seen a few cases of people with amazing pieces of hardware mixed into a case that was so imbalanced as to make you cringe.

Why would anybody mix a pair of SLi 7800GTX video cards with an under-par power source or put an AMD 4000+ X2 processor into a cheap option-free mainboard?

All I can see is ignorance. Not to say it's anybody's fault, but ignorance when building yourself a high-end (or so you think) system can cost you more than the obvious!

In this particular case, running a sub-par power source with such high-consumption hardware as 7800GTX's and SLi can cause catastrophic system failures. When the power supply is over-taxed, then it can start dropping rails and fluctuating voltages.

Somewhat akin to a brown-out of your hardware, anything from lost data to completely damaged hardware can occur.

In the case of the CPU and mainboard mis-match, the consumer should have saved a few $$$ on a slightly less expensive processor and upgraded the mainboard to a high-performance version (Think DFI Lanparty SLi or MSI K8N Diamond).

Of course, in the same mixture is memory! Especially on AMD machines, running slow high-latency memory modules with an extremely demanding processor (and adequate mainboard) can end up losing you quite a lot of performance, especially in applications requiring lots of random accesses. Applications like games.

What I'm getting at here is that if you plan on a supercomputer with the very best processor, but don't plan for the rest of your hardware, you will end up wasting a lot of the power of that piece.

When it get's down to it, the processing performance of your system is directly related to the performance of your processor, mainboard, memory and, believe it or not, your power supply. Remember what fluctuating voltages do to a system? Yeah, at full speed these components can use a lot of Wattage and you want to make sure the PSU you've chosen can supply it.

When we expand our view to include graphical prowess, you must also include the video card. Since the video card connects through your mainboard's chipset, you can easily see how important a high-quality mainboard can be.

Even if two boards have the same chipset on them, their performance can vary dramatically due to it's implementation. One poor trace route or bad shield can result in a poor signal giving you anything from minor artifacts to full-out crashes. Poor routing and implementation is also a big contributor to poor performance.

FYI, many of the cheaper mainboards (this isn't just limited to mainboards) use cheaper and less reliable components to keep down costs. While these may work fine for stock settings, the so called extras of the board may be sub-par or unstable and don't count on any overclocking ability. With electronics, you usually pay for what you get. Granted, there are some names (Read: ASUS) that you will pay a premium for, but their high-end revisions, usually named Premium, have the best components available as a company tradition!

What's all this getting to? Please don't get your wallet stuck on something you can't actually afford at the cost of using cheap or insuffiicient hardware elsewhere in your system. A lower CPU and better board would have done a better job and yielded a higher performance than the best CPU and a nearly worthless mainboard.

Make sure your systems are balanced when you buy them. If you don't know, that's why forums like these are in existence. Just ask.

PLUR
CK

[CandyKid]

Here's another one...

Why should I spend more cash on a power supply?

This is easy: Because every piece of hardware in your entire system is connected to it. If the power supply blows up, it can destroy anything and everything connected to it.

Cheap power supplies usually use inferior components (havn't we heard this before) and will not be nearly as reliable as many of the better brands.

Insufficient power can, as said above, seriously damage your hardware and cause instability, rebooting, inability to overclock, loss of performance and many, many other problems.

Now, you might ask, how can I tell a cheap power supply?

Any power supply that is extremely lightweight, I'd shy away from. The outer construction should feel very solid and not flimsy at all and the innerds shouldn't be loose.

One of the best ways to tell a good power supply is the MOSFET heat-sinks. Inside almost all PSU's will be two heat sinks, most look like the letter "T" comming up from the circuit board. If these are large, heavy, and highly-folded, then they are high-quality.

If the heat sinks are very thin, do not have a "T" top to them or only a few folds/fins, then they are much lower quality and will probably not adequatly cool your power circuitry.

In my own experience, this resulted in a burnt circuit board and a few fried resistors. The PSU lasted for about six months, taking with it multiple motherboards!

Basically, a good power supply will feel very solid, have some weight to it, and have a good cooling solution. This includes the fan and the heat-sinks, especially.

So, another question is why shouldn't I use the power supply that came with my case?

Because 99% of them are crap. Sorry. :p

But seriously, most of the OEM PSU's are won from a "who's the cheapest" bidding war and are very low quality. You'll hear stories of how many people used them for years without a problem, but it's not the norm, especially for demanding high-performance systems. Keeping the OEM PSU is a risky bit of business which I would avoid.

HOWEVER, there are a few cases that come with a high-quality PSU... you'll pay dearly for that case, but it can be had. Just don't expect the $40 USD case + 400 Watt PSU to perform worth a damn. Buy a better PSU instead and get a case without one.

By better brands, I mean PC Power & Cooling. They are the industry power supply god and are expensive as hell, but I've never seen one of them fail to outperform any other similar model on the market. They're built like tanks and last forever.

Somewhere below PCP&P, but still VERY high quality models are made by Fortron/FSP, Sparkle, Enermax, OCZ, Seasonic, Tagan, and Zalman. This isn't a complete list, but should get you started.

Some brands, I.E. Antec, have a huge fan-base and lots of advertising, but their products are going downhill. Others I'm going to get shot for this like Thermaltake seem to be going up in quality, but they're still on a knife's edge. Some are good, some are bad.

Brands to avoid would include Aspire, Chieftech, Kingwin, NEC, PowerStream, Tyan, Ultra, Zippy, Powmax and any OEM manufacturer like Dell, Gateway, Hewlett Packard and the like.

Regardless, don't jeopardize your system by skimping out on the most overlooked piece of hardware in your system. Buy a good power supply to create a solid foundation for the rest of your system.

PLUR
CK

p.s. Special thanks to seamaiden. They know why.

[Pug]

By better brands, I mean PC Power & Cooling. They are the industry power supply god and are expensive as hell, but I've never seen one of them fail to outperform any other similar model on the market. They're built like tanks and last forever.
I borked my PCP&C. :o)

I'd like to add HeroIchi (HEC) to the list of tried & tested - they can often have budget fans in the low-end models (read noisy but easier to mod than other cost-cutting skimps). Engelking are very good (and do a lovely SFF PSU), Seasonic have likewise produced some nice bits o' kit.

On the mid-level, my eye is currently on Hiper (hey, I'm on a budget too) on a par with ThermalTake in the 'are you sure you don't have another tenner in the budget?' category but I have limited experierence with these, as yet (subject to change).

Adding to the list to avoid; Deer, Q-Tec([edit]not to be confused with hnology where applicable), and many others that escape me atm.

Kev and I had an in-depth discussion, a few years back, on the joys of Active PFC (Power Factor Correction).
I've since found an awesome link, so I'll shut up and let you read some. :-)
http://www.efficientpowersupplies.org/

[Greeny]

Q-technology is not the same as Q-tec iirc, Q-techologies PSU's are better quality though I have never owned or used one I've seen a few reviews and they look to be of reasonably good quality.

http://www.qtec.info/ - crap
http://www.qtechnology.net - look half decent

[CandyKid]

Interesting stuff guys, I'd never heard of Q-Tec/hnology over here in the states!

Thermaltake, however, has impressed me with their ATX 2.0 PSU's. Mine is solid as a rock and I really do enjoy it, even when overclocking it's rails are near dead-on. Definately within spec.

Don't read too much there, though, a lot of their old stuff is crap and all bling-bling.

Oh well, too many people are getting into the PSU business, IMHO!

PLUR
CK

You broke your PCP&P? Wow. That's just insanity man, what the heck did you do to it!? FYI, there was a slick review of some PSU's on Tom's which tested the PCP&P 850 SSI. What a monster. Go try to kill that.

[CandyKid]

Here's one for you that's very near and dear to my heart as whenever I remember to ship my modules for RMA, I'll be trying the same thing.

I bought two pairs of matched dual-channel memory modules rated at 3-2-2-8 and 675MHz, but they won't run at their rated specs! Why not?

Memory modules are rated in dual-packs because anything above two modules is almost impossible to match.

The possibilities are so numerous when you take into account the plethora of chipsets and computer setups that it gets to be overwhelming.

But if I have four identical memory modules and my mainboard says it supports them, why won't they run at their specified speeds and latencies?

As stated elsewhere, more modules means more loading on the memory chipset which means an access takes more time... if an access takes more time, then fast/rated latencies may be too quick for the system to keep up with and may need to be relaxed.

So, while two modules are rated at 3-2-2-8, your system may only be able to run four modules at 4-3-3-12 because it's basically 'too much' for the system to keep up with.

The other limitation is that all of the modules must act the same way for the best performance and best chance of obtaining lower latencies... if one module acts differently, then the whole system must slow down to the weakest point.

Say, for instance, you have a CAS 3 part and a CAS 4 part. The system will automatically run them all at CAS 4 as it cannot run them at different speeds and cannot run the CAS 4 part at CAS 3.

That's why latencies must change, but what about the speed?

Just had a guy complain about PC4000 modules in an AMD system and wondered why they wouldn't hack 250MHz.

This goes back to remembering that the rest of your hardware makes a huge difference on your overclocking ability.

Memory rated at PC4000 is not automatically run at 250MHz, it's run at whatever default speed your system has set in the BIOS! If you're CPU isn't overclocked and you set the memory to a 1:1 divider, it'll run it at 200MHz/PC3200 speeds. The PC4000 denomination only means that it can run at 250MHz, not that your system can.

Just a couple of make-my-life-easier memory pointers:

1) Plan ahead. If you have the ability, match your memory modules now and get enough to last you awhile. Adding modules later and/or mixing and matching can be tricky and lose you performance.

2) Don't go overboard. If you don't need all four memory slots filled, then don't do it. There's a big difference between wanting 4GB of RAM and needing 4GB of RAM. FYI, Windows XP won't even cache anything more than 2GB of RAM for program use.

3) Research your board and CPU. Don't expect a locked AMD 1900+ on a 266 FSB to run PC4000 memory 1:1 at 250MHz. That'd be a pretty big overclock and most nForce Ultra 400 chipsets won't handle it very well, much less this CPU. Know your hardware.

4) Do you need faster memory or more of it? Planning on playing games? Then buying 1GB of value-RAM is probably a better idea than buying 512MB of expensive ultra-fast-tight-timings memory.

5) Remember, memory can be as inexpensive or as expensive as you want it to be, but buying brand-name warranty-backed modules is always a good idea.

PLUR
CK

[CandyKid]

Here's one that needed following up!

Original Question: Why is the ASUS 6600GT TOP Limited Edition Special!?

The ASUS 6600GT TOP Limited Edition video card is special not because it has a new core or different fabrication process (although a 90nm GPU would rock), it is something special because of it's stricter quality specifications and high-end components.

Most companies, BFG and other mainstream overclocked card-makers included, don't use the absolute best components available. Why spend an extra 3 cents per card when you don't have to, especially when thousands are being made because that adds up fast!

I'm not saying there's anything wrong with that either. You don't neet a Porsche to get you to the grocery store, a Volvo can still do it in style.

However, the TOP Edition does use some of the best components available including the 1.6ns GDDR3 chips (yes, Samsung) and some very nice capacitors. The whole design uses very high-quality components and even the heat-sink has something a little special (note the color difference between the TOP version and non-TOP versions).

What this means is that the card should reliably overclock higher and be more stable than a competitors card (or even the non-TOP ASUS cards), but will make them a bit more expensive.

The reason only 5,000 are being made is because there's not much of a demand for ultra-quality boards when there's so many other high-quality versions available.

Unfortunately, the only issue with the card is an important one: The thermal compound isn't very good compared to the rest of the card! While the original spec was to release the card at 550/1100, it was found that it created too much heat and instability and thus you can buy the card stock at 520/1100.

The good news is that this can be easily remedied with a bit of isopropyl alcohol and Arctic Silver V or Arctic Ceramique. Just be careful not to torque the edges of the GPU when removing/installing the heat sink. Quite a few users have discovered that this simple solution allows fairly consistent overclocking past the original 550 on the core. I have one and with this fix made it quite a ways past that while remaining stable, but it get's a bit warm (soon to be water cooled, so we'll have to see).


ASUS has released a 7800GTX TOP Edition!!!

Wow, what a monster. Starts at 486/1350 Core/Memory and has been overclocked upwards of 520MHz on the core. Amazing. Looks to use the same 1.6ns GDDR3 as the 6600GT TOP did!

Also uses the nV Silencer 5 as it's thermal solution (rebranded) so it's a two-slot fix, but runs ~20 Celcius cooler than the non-overclocked reference design!

I've seen this card for $550.00 USD, so it's not astronomical compared to the other 7800GTX's out there, but it packs almost 12% performance over the rest! WOW!

In lieu of the 7800 ULTRA cancellation, this seems like a great choice for that last bit of performance. We'll still have to see if ATi's r520 can compare to the G70 series, but, more importantly, can they make up for so much lost time!?

PLUR
CK

http://www.asus.com/products4.aspx?l1=2&l2=6&l3=0&model=589&modelmenu=1

[CandyKid]

I got this one via PM:

I didn't want to waste space on the forum by asking a question about this but since you posted it I'll ask you. in the DDR-II thread you said normal ddr memory runs where every signal there's 2 bits of information sent, so with DDR-II it's been changed to four, BUT it runs slower?

That is correct. DDR-II can send the same VOLUME of data bits at 100MHz as DDR-II can send at 200MHz, but they'd BOTH get the value of PC3200! DDR-II = 4 x 100MHz x 8bits = 3200 while DDR-I = 2 x 200MHz x 8bits = 3200.

I'm confused just curious but wouldn't thta make it a lot more able to be OC'd? Also I'm assuming the speed of RAM is the internal speed of the memory, or is the speed the information is released from a stick of RAM.

Yes and no... first off, PC8000 / DDR-II 1000 is truly clocked at 250MHz which is the same as PC4000 for DDR-I. The exact same real clock, but it's sending twice the data and so you get twice the bandwidth! This also means, however, that the chipset and input/output registers are running twice as fast and we can quickly see how much of an 'overclock' that turns out to be.

So yes, DDR-II can go 'faster' but perhaps not in the same way you're thinking and also as far as external clocks go.

Also, the internal/external speed of the RAM is not the same thing. What speed is posted is usually in two parts...

PC8000 is the nomenclature used to denote the theoretical bandwidth and because it's usually a big number, it looks good.

DDR-II 1000 is the nomenclature used to denote the effective clock speed of data.

The one they never tell you anymore is the ACTUAL clock speed, but because of the way the memory works, it doesn't look good to pay TONS for 200MHz DDR-II when you can get 275MHz DDR-I for a fraction of the cost (PC4400).

Lastly you said that number (PC3200, PC2700, etc.) is just actually one half of what the speed of the RAM is? so PC3200 runs at 200mhz and sens 2 bits each time so they say it's clocked at 400? or do they give the 400 number to show how much information is sent within those 200 mhz?

Err... read above again, I believe you mis-wrote this question.

However, the 400 denotes the data rate for DDR-I at 200MHz. DDR-II at 200MHz is 800 or PC6400 compared to PC3200 for DDR-I.

PLUR
CK

[CandyKid]

Something simple, but more personally involving:

I bought PC5400 / DDR-II 667 memory modules, but the 'tested' settings have the frequency at 533!

I just figured this one out, but it took my dense head awhile. What we do here is remember our frequencies!

A 200MHz FSB with a 3:2 divider yields a 133MHz memory frequency times four for DDR-II equals DDR-II 533 / PC4300!

Now, if you OC your FSB to 250... a 3:2 divider yields a 166MHz memory frequency times four equals DDR-II 667 / PC5400!

So, by saying they set it to 533, what they really mean is they're setting the memory divider, then increasing it by overclocking the CPU's FSB (thus increasing the memory's frequency).

Personally, however, you should also be aware of the optimizations of your specific chipset. The i915's and 925's are optimized to run a divider of 2:3 or 4:5, but not 1:1 at high frequencies. The i945's and i955's have been changed for optimal performance 1:1 with the FSB at high frequencies, hence their support for DDR-II 800 out of the box.

PLUR
CK

[CandyKid]

What about optimizations for the i875 and 865?

AFAIK, they were designed with PC3200 memory in mind, I.E. 1:1... that's why using dividers on the i925 boards seems so much easier than it used to be, they're meant for it and handle large dividers (1:2 or 2:3) very well, whereas the older chipsets were touch and go on that sort of thing.

Another consideration along these lines is that with the i865 and i875 chipsets, running DDR400 with a 533 FSB was impossible as Intel left out such support... mostly because it's not very stable!

PLUR
CK

[CandyKid]

What memory speed/type is right for me?

Matching your modules to the needs of your system and your performance expectations can be a tricky matter, but doesn't have to be!

First of all, you should decide on HOW MUCH memory you're going to need and whether or not it is possible to get it at one time, or if you'll need to buy some now, buy some later. With 32-bit platforms, Windows XP will not cache over 2GB's of RAM for program use. Depending on your setup, the rest can be used by the page file, but is mostly useless. 64-bit systems solve this limitation, but in the end having 4GB of RAM usually slows your computer down more than speeds it up.

Only buy what you'll reasonably need which for a long time has been 1GB of RAM, preferrably in a 2 x 512MB dual-channel kit. There's just starting to be programs that actually see a performance boost from having more memory than that, most notable is Battlefield 2. Of course, there's always a couple of programs (Photoshoppery) that can use more than 1GB of memory, but for most users 1GB is enough.

Next you'll need to decide on the speed and type of the modules... PC1066, PC3200, PC4400, PC4300... they all look alike, but that list actually has three different types of memory in the list! Almost all of today's computers will run with either DDR SDRAM or DDR-II SDRAM memory modules. While any mainboard specification will tell you what type the board supports, here's a quick rundown based on the type of CPU:

AMD Athlon XP and other Socket 462 systems support DDR memory modules (you can find a rare one that supports SDR modules, but it's not worth your money)

AMD Athlon 64 Socket 754 and Socket 939 systems support DDR memory modules while the Socket 940 version requires a special function of memory and modules that are registered. Unless specified, un-buffered modules are the norm.

Pentium 4 Socket 478 systems support DDR memory modules (again, the REALLY old versions support SDR, but still not worth your money)

Pentium 4 Socket 775 systems support both DDR and DDR-II memory modules, but only the i915 chipset and certain modified i8XX chipsets support DDR-I. All of the new chipsets support only DDR-II modules.

The server systems like Xeon or Opteron platforms all require registered and sometimes registered + ECC modules, but this is a bit more advanced than needed here.

Regardless, the most common types, by far, are DDR-I and DDR-II memory modules.

Now, down to the basics of speed!

If you're running an Intel CPU with an 800 FSB or an AMD CPU with a 400 FSB, then running it synchronously with the memory results in DDR400 or PC3200 memory modules. Let's focus on DDR-I memory modules first:

Running the memory with a 1:1 divider with the CPU's FSB results in a common frequency between the two, eliminating down-time due to mis-matched clocks and other latencies. Usually, running the memory 1:1 is the fastest and most stable of all configurations... but if you overclock, that means you'll need memory that can go FASTER than 200Mhz.

Currently, the best overclocking memory available is based on the Samsung TCCD IC and will reliably run at 275MHz. Labeled as PC4400 memory, it can usually achieve a CAS latency of 2.5 without issue. So the plan is this, if you're going to overclock, buy memory modules that are at or exceed your expected final frequency... thus the memory is taken out as a limiting factor in your overclock.

Now, it's a bit different with DDR-II because the modules can handle frequencies much higher than most CPU's FSB. For that same Intel (a socket 775 this time though) with an 800 FSB, you can pair it with DDR-II 533 / PC4300 memory modules and it doesn't even require an overclock. One of the great features of the new chipsets is their default support for very fast memory. The i915/i925 set supports PC4300 memory via Intel and most manufacturers have added support for PC4800 memory, all without overclocking your CPU. The newer i945/i955 set brings us up to PC6400 without overclocking.

Simplistically, this means that with DDR-I systems, buying PC3200 memory is the most common for stock-speed performance while faster modules are available depending on how much you want to overclock.

For DDR-II systems, I wouldn't purchase anything under PC4300 and would actually suggest getting PC5400 memory as most of the chipsets will support PC4800 speeds (there's no official DDR-II 600 modules available, so you go to the next speed up and underclock them) or faster if you buy a new chipset.

The last thing to think about is latencies. Lower latencies are better, it's that easy, but you can't always compare apples to apples! In this case, price also comes into play... 2 x 512 value modules with a latency set of 3-3-3-8 can usually be found for $100 USD while 2 x 512 high-end modules with a latency set of 2-2-2-5 can cost upwards of $200 USD or more! Depending on your use and applications run, you may or may not even see a performance difference!

So there you have it...

1) How much?

2) What kind?

3) What speed?

4) What 'quality'?

PLUR
CK

[CandyKid]

When looking at memory specifications, you'll often see a set of four or five numbers listed as their latencies. For most people, however, 2-2-2-5, 3-4-4-8, 2.5-4-4-7, 3-2-2-8 or even 5-5-5-12 probably don't mean much. Add onto the end of that a 1T or 2T specification and everybody's confused. So...

What do memory latencies mean?

While there's no set JEDEC standard for which order the latencies should be listed, the common convention is this: CAS - tRCD - tRP - tRAS with the last number being the command rate.

The CAS latency, or Column Address Strobe has several functions, but the system must wait the specified number of clock cycles every time a non-sequential read or write is initiated. Sequential read/write operations can bypass this step.

Since most of our systems have loads of random read/write operations, you can fairly easily see how this latency becomes the most important in terms of memory bandwidth. For instance, benchmarking my own system with dual-channel PC3200 memory, I gained over 400 MB/s in bandwidth going from a CAS of 3 to a CAS of 2.

The tRCD latency, or RAS to CAS Delay, is a changover between the steps of memory accessing and locating. This latency can also have a very large impact on your bandwidth! For the most part, we rarely see a tRCD below 3 and when you do, it's a very fast memory module (Think TCCD).

The tRP latency, or Row Precharge, is nearly self-explainatory! The cells are pre-charging to perform an operation. High-performance memory modules will often support a tRP of 2.

The tRAS, or Cycle Time (among 30-some other names), is a very difficult latency to explain, but lower isn't always better, TBH. Think of a memory operation like reading a book... you have to find the page that its on, then you have to find the line that you want (Think RAS & CAS), but once you've got it, you have to actually read what it says... think of the tRAS as how long you have to do that before the book is slammed shut in your face! Technically speaking, half of the word must have been sent and the other half in the output registers by this time, but that's just from memory and may not be completely correct.

The Command Rate setting defines how many clock cycles it takes for the CS (Chip Select) command to latch a signal onto the DRAM.

In English, that means that a 1T Command Rate requires one clock cycle for the chipset/MCH (memory controller hub) to select which bank/chip the memory access commands are going to. If you have only one single-rank memory module, this process is simple. If you have four double-rank memory modules, then this process is complex and will probably require two clock cycles, 2T.

The Command Rate is more hardware dependant and rarely memory dependant because it isn't an internal memory access latency (I.E. CAS) and doesn't require the memory to actually do anything, it's the MCH that's working at this point.

Lowering your latencies is always a good way to optimize your memory's functions, but often times also requires that you give them more voltage. To this end, with DDR-I, I'd never exceed 2.9 volts and with DDR-II, I'd never go past 2.1 volts.

PLUR
CK

p.s. FYI, DDR-II modules pull only about 7 Watts apiece while DDR-I modules pull about 12 Watts apiece!

[CandyKid]

Apparently I said, "Random single-bit errors are usually caused by low voltage or incorrect latencies," around here somewhere and here's why!

Low voltages can actually imitate incorrect latencies and both end up doing the same thing... If a high signal is a 1 and a low signal is a 0, then a non-signal (I.E. not enough voltage to trip the transistor) would also be read as a 0.

So if your memory doesn't have a strong signal, then every time it is read from, there's a chance that one of the signals won't be strong enough to signal a 1 and will incorrectly be read as a 0.

The same goes for latencies, if a latency is set too aggressively, then the system may not have time to fully charge a cell before the signal is sent, resulting in a weak signal...

These errors occur randomly and every read has the same chance of being a weak signal. The more deviation is experience, the more errors will occur (sometimes thousands or more), but they won't be repeatable.

If you ran Memtest86+ (http://www.memtest.org/) and the same location caused an error repeatedly, then that would indicate that the memory itself was faulty and that particular physical address had something wrong with it.

Also, if you've double, triple, and quadruple checked your settings or tried less-aggressive settings and a higher VDimm (memory voltage BIOS setting) and still get loads of errors, then that as well would indicate a faulty module, but a different part.

PLUR
CK

p.s. Low enough voltages can be had that the memory cannot effectively refresh and thus loses massive amounts of data, but at this point, your system wouldn't even boot.

[Pug]

I don't want to disrupt the flow of this awesome mine of information but I can't help myself instigating a quick standing ovation here. :applaud:

Damn good stuff mate, all I'd forgotten and more. :bow:

[CandyKid]

...dangit, anybody that stands up has to ask a question!

That means you too Pug!

PLUR
CK

[CandyKid]

CK could you enlighten us with a more indepth look at memory dividers? Its mentioned a lot in your posts and is often essential in OC'ing so an idiots guide would be good.

Ditto. I'd never heard of it before until I started reading up on OCing the Skt 939 Opterons, I think I have a basic idea of what it is but I'm unsure as to it's effects in practice.

its just a ratio of FSB frequency (or hypertransport for you amd64'bitters) to memory frequency. Thats all a divider is, their is nothing complicated to it, grab a calculator and have a play. Let's say you have a 3:4 ratio memory divider that would mean that for every 3Mhz you have in FSB your memory speed will be 4Mhz, e.g. 3:4 at 100Mhz FSB your FSB will be 100Mhz and your memory speed will be 100/3 x 4 = 133Mhz.

That's exactly correct. Moreover, you need to be careful when selecting your frequencies because Intel systems often list only a frequency (and not a divider) and AMD systems give you TWO multipliers.

On the Intel side, especially with DDR-II, the dividers become more complicated as you can set the memory frequency higher than the FSB frequency... sort of.

First off, you need to know how to calculate memory speeds. This is different for DDR-I and DDR-II, here you go:

FSB Frequency * Divider * 2-bit sent per cycle * 8-bit DWORD = PCXXXX DDR-I

FSB Frequency * Divider * 4-bit sent per cycle * 8-bit DWORD = PCXXXX DDR-II

What I mean by 'sort of' is that you can run DDR-II 800 with an 800FSB CPU, normally running 1:1 with PC3200, now running with PC6400. This would seem like a 1 to 2 divider (I.E. 200MHz BUS / 1 * 2 = 400MHz * 2-bit/cycle * 8-bit = PC6400), but it isn't. It's actually a 1:1 divider yet, but the math is a little different due to DDR-II's ability to send 4 bits per cycle as discussed earlier (Thus 200MHz BUS / 1 * 1 = 200MHz * 4-bit/cycle * 8-bit = PC6400).

On AMD's, you have the option to set both a memory divider and an HTT divider. Let's just get rid of the 1000MHz myth and admit that the FSB runs on a 200MHz frequency. The memory controller AND the CPU run off this BUS.

Current CPU's boast a 1GHz BUS which is a 200MHz frequency * a 5x HTT. The older A64 CPU's had an 800MHz BUS which was actually a 4x HTT.

Unlike in Greeny's example, current Intel and AMD systems don't allow you to set a divider higher than 1 to 1. This means that to use anything above DDR-I PC3200 or DDR-II PC6400 requires you to overclock your system. Older systems with slower FSB's (any frequency below 200MHz) will allow you to run the memory faster than the FSB, but it's often a bad idea for stability.

On a side note, i915 and 925X based systems don't allow for a 1:1 memory divider (DDR-II 800) and only allow up to a 5:4 (DDR-II 667) divider. On the i925XE chipsets with FSB1066 support, the 3:2 divider is what's being used to support DDR-II 711 when using a P4EE CPU.

So, overclocking... gimmie a minute on this one, boss-man is wondering why I'm not working hard today... damn hump-day blues. I'll edit this post and add in more goodies sometime today.

PLUR
CK

[jaguarking11]

Another consideration along these lines is that with the i865 and i875 chipsets, running DDR400 with a 533 FSB was impossible as Intel left out such support... mostly because it's not very stable!

Actually my ASUS P5P800 with an i865PE chipset can run the memory higher than the FSB by setting it to 533. I don't know if ASUS screwed up and adapted an i915 BIOS or they went nuts, but with that sort of divider I was able to run up to DDR520 with Kingmax Hardcore DDR-1. I checked it with CPU-Z (www.cpuid.com) and the memory is definitely running faster... but it did not net any performance increase. So I guess your right, the chipsets were not designed for it.

[CandyKid]

It would seem that they have, in fact, ported some of the newer architecture and/or BIOS into this rehash board. Apparently, the implementation was a good one judging by your clocks!

PLUR
CK

[CandyKid]

CK my Graphics card uses GDR3 memory? I know this is more advanced than DDR2 but how does it differ?

Also how is its speed rated? I have mine overclocked via ATItool to 620mhz so if i was marketing guy how would i describe that speed? E.G. 200MHz DDR RAM is marketed as DDR400. Does GDR3 memory use a similar convention?

Err... Read This (http://www.lostcircuits.com/memory/ddr3/).

As far as I can tell, the speed ratings of DDR3 and DDR2 will be the same, 4-bits per core clock... as already discussed, however, these are not the same as DDR1...

I.E. DDR1 @ 200MHz is DDR400, but DDR2 @ 200MHz is DDR800. Regardless, most marketing has already tried to eliminate the difference by rating 200MHz DDR2 modules as 400MHz which in the old naming scheme would be DDR800 and make more sense to the layman.

The 1GHz memory you're talking about depends on the card, but a TRUE 500MHz is possible when the DRAM core is back up to 2.5 volts while leaving the interface at 1.8 volts. Of course, this creates a LOT of heat, but it can be done.

I'd be more likely to assume that a 1GHz rating was the result of a true 250MHz clock which, for DDR2/3 would be DDR1000... tossing numbers that large almost always means data rate rather than clock speed.

I was wondering if it was different types of RAM or just different ways of listing the speed.

Point blank, DDR3 is a different architecture (and physically different memory, DDR3 for the system will have a different pin-out footprint) than DDR2, but utilizes the same BUS.

In a nutshell, GDDR3 has a couple of innovative features such as noise-reducing architecture allowing for higher clock speeds and GDDR3 has a higher refresh rate than DDR-2. A source more knowledgeable than myself said DDR3 can go upwards of 1600MHz (note, DDR1 has proven to go WAY past it's original speed spec, so we'll have to see how fast DDR3 can go).

PLUR
CK

p.s. Sorry for the lax answers of late, been busy and stressed!

[CandyKid]

i have always come too believe that a single larger strip of memory is better than having two smaller sticks. why is it better too have 1gb but in a 2x 512mb sticks.

Stolen from here (http://www.houseofhelp.com/v3/showthread.php?t=24552).

"Dual-Channel Architecture" is a physical property of the memory controller (on chip or otherwise), not the memory.

A normal Single-Channel motherboard has one 64-bit memory controller and thus one 64-bit interface. Each module of DDR also has a 64-bit interface (unless it's ECC then you get an extra 8 bits = 72-bit). This single controller accesses memory modules in all of the DIMM slots.

A Dual-Channel motherboard has two 64-bit memory controllers and thus one 128-bit interface. Often you'll see something advertising 128-bit memory, etc. This is the basis since Dual-Channel uses the same 64-bit DDR memory.

However, in Dual-Channel boards, the first memory controller only accesses modules from two DIMM slots. The second controller accesses modules from the other two so you have two controllers simultaneously accessing memory.

Theoretically, this doubles the speed at which all memory related functions can occur. In real life, it's not doubled because of physical restrictions and other interactions, but it usually does give a nice boost in memory performance.

You may have noticed that neither controller accesses memory from the other's respective DIMM slots. This means you must use at least two modules of memory in order to run a Dual-Channel setup. One module for each controller to access. If you only run one module, even in a Dual-Channel capable motherboard, you'll run it as if it were a Single-Channel motherboard since only one of the controllers has access to it.

The underlying idea is that when one controller is working the other is resetting or preparing to access/write memory and vice-versa so it reduces memory, "down time."

The biggest requirement is that the modules you use are closely enough matched that they can run almost exactly the same. They will run with the same latency settings, same speed, etc. For example, Corsair tests their TwinX sets in a Dual-Channel configuration to make sure they can, "take the heat," but if you bought their Value Select memory, they do not test it in Dual-Channel. However, many times people are still able to run them as such. You have a better chance of it working the more high quality the RAM is (to simplify things).

In the real world, you may or may not ever notice a difference in performance. It depends on the types of programs you run and how memory dependant they are. Many video games would probably get a boost, but nothing too substantial because of their dependance on other hardware such as the video card. The more your programs need to access RAM, the more of a performance boost you'll recieve.


also about matching memory is it correct the computer will work better if the memory is of the same make and style of chip layout. i was told that the computer will always accses the first bank of memory it sees, in mine that would be dimm socket 1 256mb un-named and my second dimm slot 2 would be 512mb samsung.

I'm not sure I know what you mean, but the memory controller reads the SPD chip off of every module and will automatically set the highest common denominator for latencies and frequency. Also, as said above, yes, the system will be more stable and more overclockable when the modules are matched.

would i be better of with 2 x 512mb or 1gb and 1x 512mb. i cant afford 2gb but 1.5gb is in my range.

You'd be better of with 2 x 1GB or 2 x 512MB which evens out the load on each memory controller and makes the process of running dual-channel easier. I am assuming here that you're going to run a dual-channel system.

1 more thing sorry. is it correct the more money memory is the better quality you are recieving.

Not always. Look for two things: Quality and Support. I can only truly vouch for Corsair, but OCZ makes some good modules along with Patriot (Patriot seems to make theirs a bit less expensive as well).

PLUR
CK

[CandyKid]

thank you CK i now understand how it works roughly. i will look for a dual board now

FWIW...

Most nVidia nForce 2 chipset implementations are dual-channel.

Most Intel i865 and i875 chipset implementations are dual-channel.

Some SiS chipsets such as the SiS655 (Intel s478) are dual-channel.

AMD Athlon 64 s754 CPU's are NOT dual-channel.

AMD Athlon 64 s939/940 and Intel LGA775 based systems are all dual-channel.

PLUR
CK

[CandyKid]

CK, a question, im my bois i have a burst length option (either 2, 4 or 8 bits) whats this?

The burst length is how many bits are transmitted in any 'send' cycle (not the same as a clock cycle). 8 is optimum, but sometimes impossible with older chipsets, thus 4 is also fairly common. 2 would just be a waste of performance.

also i cant seem to find my ram divider, so my ram (ddr400) is running at 250 (is this not ddr500 speeds?) i dont think my bios actually has one. haha.

Technically, you do have a RAM divider, but not everybody displays it as a X:Y divider. Often, they just list the resulting speed of the RAM, but if you do the math, you can determine the divider.

Also, many systems allow you to adjust your RAM frequency in 1 MHz steps... sort of 'divider-free' if you will, but in addition, they'll still have several pre-set choices.

If it's running at 250MHz, then yes, that's DDR500 and many DDR400 IC's will hit it easily (think TCCD among others).

PLUR
CK

[CandyKid]

okay ram man... I got one for you...

4 x 512MB of PC4000 ballistix vs 2x1GB of corsair XMS PC3200 (I don't have the model #s off hand, sorry.)

This depends. If you're running the PC4000 at 250MHz, the Corsair modules probably will not reach that speed.

If you're running the PC4000 modules at PC3200, then look to the latencies for a comparison.

1) If one has lower latencies, that's a plus.

2) If you cannot run the 4 x 512 at a 1T command rate, I'd use the 2 x 1GB.

3) A 1T command rate is MORE IMPORTANT than lower latencies. CAS 2 @ 2T is worse than CAS 3 @ 1T.

Are the PC3200s BH5 you think? the stuff is probably ~2003 or 2002 for manufactur.

Not a chance. Winbond -B Rev 5 IC's are 32 Mbit IC's which cannot possibly make a 1GB module using conventional manufacturing techniques. 16 of these IC's (8 per side) results in a 512MB module.

I never should have box swapped it, and am thinking of swapping my 4x512 into these (non overclocked) boxen in favor of the 2x1GB sticks. The 1GB sticks aren't low latency stuff, either

Since you'd be running both setups at PC3200 speeds, the fact that the Ballistix are PC4000 is moot. So compare the latencies you're able to run, including the command rate.

Also, make sure to check that they both run in dual-channel (if your rig is capable) and, for that final push in either direction, run Everest (http://www.majorgeeks.com/download4181.html) and check your memory latency (access latency in nanoseconds) along with the read and write performance of each setup.

That ought to give you the answer you're looking for.

PLUR
CK

p.s. If you give me the full model name, revision number, and lot code of your Corsair modules, I'll tell you anything you want to know about them.

[Starbuck3733T]

You're a god amongst men. the IC7 I had couldn't push the ballistx in 1T command rate, but I was pushing ~270FSB through them at one point. Dual channel for sure. PAT said enabled via CPUZ. I don't know about the Asus rig and being able to do 1T through 4 sticks.

I asked about the BH5'd ness of them because I know the micron 5B chips that the ballistix use don't like extra voltage to get low latencies the way the BH5 stuff does. I'll TRY to get the model #s for you.

[CandyKid]

What Programs Can I Use To Test My Memory/System/Graphics?

CPU-Z (www.cpuid.com/cpuz) - Simple, clean, useful. Information about chipsets, current clock frequencies, and memory SPD settings.

Motherboard Monitor (MBM) (http://mbm.livewiredev.com/download.html) - Monitors temperatures, voltages, fan RPM's, etc.

SpeedFan (http://www.almico.com/sfdownload.php) - Same as MBM, but kept more up to date and has more support.

Everest Free Edition (http://www.majorgeeks.com/download4181.html) - Huge amounts of information at your fingertips about most every piece of your computer plus a very comprehensive memory read, write, and latency benchmark.

Sisoft's Sandra Lite (http://www.sisoftware.co.uk/index.html?dir=dload&location=sware_dl_3264&langx=en&a=) - Didn't know you could benchmark your network card? Sandra lets you monitor, benchmark, and investigate more pieces of hardware in your system than I've ever heard about. Great program. Also, the benchmarks are extremely useful in accessesing actual performance changes as they run using all the available resources and bufering of your system.

PC Mark & 3D Mark (http://www.futuremark.com/download/) - The industry standard tests for graphical prowess and very good stress-tests for your overall system. PRO versions are for sale, but free versions are more than adequate.

Aquamark 3 (http://downloads.guru3d.com/download.php?det=673) - An older GFX benchmark, but still useful.

Super Pi (http://files.extremeoverclocking.com/file.php?f=36) - Used to test CPU abilities and is fairly sensitive to CPU frequency changes.

PLUR
CK

[CandyKid]

Conroe and AM2 Cometh!

The new Intel Core and AMD's latest socket, AM2, will both have official support for STOCK DDR-2 800/PC6400. This means that the overclockers among us should be looking at DDR-2 1000/PC8000 or the newer DDR-2 1066/8500.

Seeing as the new 'Fat Body' chips recently came onto the market which have been showing great promise for lower latencies at insane clock speeds, these are THE modules to have. What I'm seeing is DDR-2 1000 modules runnning CAS 4-5-4- latencies and, to be honest, that's amazing especially considering these are 1GB modules. Oh yeah, a 2GB kit running at 1000MHz with these latencies? Anybody else want to buy a new computer?

Have fun with those OC's and remember, with the ultra-low heat specs on the new Conroe CPU's, I'm expecting a lot of action out of them!

FAT BODY CHIPS ARE NOT NEW

It's come to my attention that the 'Fat Body' IC's that were making the news are not, in fact anything new. They are the same chips that came out in Corsair's PC8000UL modules and other companies are just picking up some leftovers. Apparently, the media blitz is more of a selling stunt since there's not enough of these IC's left to make large batches anymore. Great IC's, for sure, but not worth a bank-breaking price as something new is always around the corner.

PLUR
CK