RADIFIED
Guide to the
Intel Northwood Pentium 4 CPU

P
A
G
E

T
W
O


O
F

T
H
R
E
E

This page continues from Page 1 of the Guide to the [Intel Northwood Pentium 4 CPU].

Until recently, the only way to run an Intel Pentium 4 CPU was with Rambus RAM. The release of Intel's i845D chipset, which supports DDR RAM (Double Data Rate), has changed this. The problem with Rambus RAM used to be that it was significantly more expen$ive than DDR RAM, without providing a corresponding increase in performance. But all that has changed in recent weeks. Prices are now roughly equal.

It's rumored that Intel cut a deal with Rambus, which awarded Intel mega-buck$ if they helped Rambus sell a lot of their RAM (by manufacturing chipsets that require Rambus RAM). I haven't researched these rumors enough to know if they are true, but I do know many people dislike Rambus. For these people, it's a matter of principle not to buy Rambus RAM.

At least one article suggests that Rambus "took unfair advantage of open JEDEC discussions, and patented concepts developed in industry committees." And the courts have recently found Rambus guilty of fraud. Personally, I'm too much of a capitalist to care about the details of techno-ethics. I merely want maximum performance for my buck, with rock-solid stability.

The price of Rambus RAM has dropped dramatically over the last few months, but is still significantly more expensive than DDR RAM, which performs comparatively to (and perhaps even better than) Rambus RAM, due to its lower latency.

You can get an idea of the relative prices of both Rambus & DDR RAM at either Mushkin or Crucial. At time of this writing, a single 256MB stick of non-ECC Mushkin 800MHz Rambus RAM cost $149. Crucial sells its 256MB stick of PC2100 DDR RAM for $88 (free shipping).

I searched Pricewatch and found 256MB sticks of PC800 Samsung Rambus RAM for $95 at New Egg (Cosair) .. better than Mushkin's $129. Others had better prices, but I refuse to purchase anything from a reseller with a poor resellerrating. New Egg's reseller rating is posted here. RJtech had some nice prices, too. Googlegear has Samsung PC800 here.

The best overclocking RDRAM RIMMs are reportedly the single-density, double-sided Samsung modules. You want what is called 16 device (low density) memory for 256megs (8 device for 128megs). High density 8-device 256MB modules (4 device 128MB modules) do not overclock as well. You want the kind that have chips on both sides of the module, not one side only. I heard that GoogleGear carries the good type, altho haven't verified this.

At least some (maybe all) of the Corsair modules use Samsung chips, but many feel that Samsung PCB is better than Corsair. (I think Samsung invented DRAM.) Newer single-sided, double-density modules do not overclock as well, altho reports are beginning to surface that some are having success overclocking the newer modules.

Note: It seems that the prices of DDR RAM have spiked in the last few weeks, while the prices of Rambus DRAM have remained constant or dropped. I heard that this is due to the increased demand caused by Intel releasing their DDR-enabled chipset. This changes the decision-making dynamics.

PC2100 transfers 2.1GB/s of data. Careful not to get GB/s confused with MHz or GHz.

P
A
G
E

T
W
O


O
F

T
H
R
E
E


Once you decide that you want a Northwood CPU, the first decision you'll have to make is: which motherboard? But before you decide "which motherboard," you need to decide whether you want to use Rambus or DDR RAM. I've researched this topic thoroughly, and still haven't come to a clear conclusion. We'll take a closer look at these questions, and find that there's no simple answer.

If you already have a bunch of DDR RAM, you'll likely want to go with a DDR-based mobo. Conversely, if you already own a few sticks of Rambus, you probably want to go with a RDRAM-based mobo. If you're like me, with nothing but regular SDRAM (PC133), the decision is more difficult. I am leaning toward going with an RDRAM solution cuz it benefits more from higher clock speeds than DDR RAM. .

The Asus P4B266 motherboard, for example, supports DDR RAM, and contains one more PCI slot (which I could use) than its Rambus-based cousin: the P4T-E. Xbit got the P4B266 to run at 172-FSB (wow). If you need USB 2.0, you could always purchase a separate PCI ad-on card for that (but it's better to save your PCI slots for things you can't get on the mobo itself).

The P4B266 comes with a 3:4 function that allows the CPU to run at 100-FSB with memory at 133-FSB .. thus maximizing memory bandwidth, without affecting the CPU. Most enthusiasts feel that this feature alone makes the P4B266 a better choice than other i845D boards.

Another favorite are the boards from MSI, which everyone seems to like. The 845 Ultra-ARU is the best DDR-based solution from MSI, cuz it offers 6 USB 2.0 connections. I've heard nothing but good things about MSI main boards, and this is the least-expensive of the DDR-based group.

The Gigabyte GA-8IRXP offers everything that the MSI boards has, but also comes with an extra PCI slot (6 total). Personally, I find the extra PCI slot an attractive feature, especially if you have lots of PCI cards like I do. Gigabyte makes quality mobos, but it is considerably more expensive than the others.

So it appears that (for DDR-based systems) the MSI board is best if price is your main concern. The Asus board is best if overclocking potential is what you're after, and Gigabyte offers the most features.


Looking at RDRAM-based solutions, the Asus P4T-E motherboard (not to be confused with the 423-pin P4T, without an 'E' on the end), will support Northwood P4 CPUs with a simple BIOS flash upgrade (see info listed under Feature Summary at bottom of the Asus page). Note that RDRAM-based systems benefits MORE from overclocking than do DDR-based systems. They ue the extra memory bandwidth better.

It contains the Intel 850 chipset (supports only Rambus DRAM). A large image of this board is posted here (122KB, 820x580). Notice how the Rambus RAM RIMM slots are perpendicular to each other. Some people feel that this perpendicular design limits overclocking potential (no hard evidence, debatable point).

Note that the older P4T-E BIOS does not recognize the Northwood CPU, and therefore can't set the clock/stepping settings. You will have to set these manually with the dipswitches, then boot, update the BIOS to a Northwood-recognizing version, shut down, and reconfigure dipswitches and jumpers back to "Jumper Free Mode".

The TH7-II from Abit uses a parallel design, which (some claim) offers superior overclocking potential. Abit motherboards are known for their overclockability, and are referred to as the 'Ferrari' of motherboards.

I heard that all TH7-II mobos manufacturered after September, 2001 sport a memory clock generator that will run/support both PC1066 and PC1200 RDRAM. I heard this from someone who claims to have contacted Abit for verification.

But Abit does not have the reputation for stability and quality that Asus boards have, which are referred to as the 'Mercedes' of motherboards. Personally I prize stability above all other factors (performance, co$t). At such high clockrates, a few hundred mega-hertz aren't going to make a whole heck of a difference. I doubt you'll notice a difference between 2.0GHz and 2.3GHz. But stability problems can make your life downright miserable.


The Holy Grail of stable overclocking is achieved when you are able to reach the next default FSB speed, with a motherboard that supports the necessary AGP and PCI dividers, so that you can run those busses at DEFAULT speeds.

Personally, I don't mind overclocking my CPUs, cuz chips come from a standard pie wafer, but I don't like running my AGP or PCI bus out of spec. This means that our CPU needs to be able to reach 533-FSB if we want to be able to run our AGP & PCI busses at DEFAULT clock speeds (provided our mobos offer the necessary AGP & PCI dividers already mentioned).

My current system (P3-700 @938MHz) is like this. The CPU is the only component being "overclocked" (which I feel is a misnomer, cuz the same pie wafer can yield chips sold at the faster speed). So was my previous system: C300a @464MHz. The only difference is that today you'll also have to run your memory out-of-spec. I've never had to do that with current/previous systems.

As Chris illustrates in his review posted here (scroll down to the table near bottom), you have the best chance of getting there with a 1.6GHz Northwood. (~US$138). He claims that your chances are "very likely". Some people claim that Northwoods stamped 'Made in Malaysia' are better than those stamped 'Made in Costa Rica', while most agree that it doesn't matter.

Overclockers.com set up up a CPU datbase query here, where you can search for typical results on any CPU you like. For the Northwood, scroll down to 'Northwood P4' near the bottom, listed under 'Select CPU type'. The average stable speed for the 1.6A (40 entries) is 2.32GHz. The average for the 1.8A (26 entries) is 2.46GHz.

I have read posts on various forums around the Net, and people seem to feel that your chances are almost certain, but you must make certain that you're getting a Northwood, and not a Williamette-based chip. A few notewothy coments from Chris' review:

A friendly tip I was given is that many 128MB RDRAM modules will overclock better than 256MB RDRAM modules, due to the 128MBit technology being more mature than the 256MBit technology, as the 256MBit is still in it's infancy.. [AND] ..

DDR SDRAM may have some compatibility issues at higher clock speeds .. [AND] ..

it's clear that RDRAM at PC-1066 speeds offers better overall performance compared to DDR400 memory paired with the Pentium 4. The dual channel design of the i850 chipset helps RDRAM gain a major advantage in memory bandwidth that simply can't be beaten .. [AND] ..

Asus's P4T-E motherboard also has these dividers, but failed to run at 533 MHz FSB due to an incompatibility with their clock generator chips.

Referring to the P4T-E's clock generator this thread contains more related info. Received a note from a reader who claims that Outside Loop is selling the P4T-E motherboard with the good clock generator .. along with pre-tested 1.6A's that are guaranteed for 2.13GHz. I would verify this with them, tho. I shot them an email and asked if they were familair with the clock generator issue. Their response:

Yes, we're familiar with the situation. Typically, when we do get the 'good' boards in, they sell out in a matter of minutes. It would be best to give us a call and place a standing order, so that you can be guaranteed to get one.

Also note that Atacom is selling boards specifically designated with the good (ICS) memory clock generator chips, which allow you to run your RDRAM at PC1066 speeds and above (if your RDRAM is good enough).

If you get a motherboard with a BIOS that does not recognize the Northwood CPU, you will have to set the dip switches to the correct multiplier, so that the board can POST. Then flash-update the BIOS so that your Northwood is recognized. Then you'll be good to go.

A word about voltage: Most people feel that increasing CPU vCore voltage 10% over manufacturers default spec is safe, but that anything more represents a gamble. If I can't get there with 15% extra, I don't go. For a chip with a default voltage of 1.50v, adding 10% would bring you to 1.65v. Note that your chip will run hotter with more voltage, so don't crank it up any higher than you need. I use Prime95 (freeware) to test CPU stability.

People are having success overclocking the Northwood P4 with the stock heatsink & fan, but if you want to take it to the next level, I've heard good things about the AVC Sunflower. Many people have said that the stock Intel heatsink installs the easiest of any heatsink they've ever used.


Update 30march2002: I couldn't resist $600 performance for a measly $134 any longer, so I upgraded my system to Northwood 1.6A (from Newegg) @ 2138 MHz, based on the Asus P4T-E (with ICS chips). See here. The most notable thing about this new system is it's stability. I've long heard from Pentium4 friends, who claim how stable their systems were. I can see now that it's true.

I was going to wait for the 815E chipset, but it is possible to run relaibly at 533-FSB now. I also liked that current board ofer 4 RIMM slots, while the new ones only offer two.

If you go this route, you need to get an Asus P4T-E with the new ICS memory clock generator chips - not the old Cypress chips. I don't know of a single person who has been able to achieve 533-FSB with the Cypress chips.

You also want to make sure that you purchase double-sided Samsund RDRAM. Googlegear specifies this. That's where I got mine. They sent me the correct RIMMs, altho not very quickly. I got 2x256MB. (512 total).

The biggest catch is that current BIOS versions (1005 right now) do not allow you to access PCI & AGP dividers to run those busses at default speeds while at 533-FSB EXCEPT by using dip-switches. But whe you use dip switches, you lose voltage control in the BIOS (jumperless). This sucks.

But you can be a bad overclocking stud by using "the wire trick" .. which UI did. It's not really that difficult. It actually requires more patience (and a good magnifying glass) than skill. If you have fat fingers, it could be a problem. More info here and here (enter as guest) and here. It works and works well. Both Motherboard Monitor and Sandra report 1.68v. I'm not sure that I actually needed the extra voltage, but merely wanted to play with the wire trick.


A quick check at Pricewatch shows that Atacom is selling the Asus P4T-E for US$155. They have *two* versions: one with audio & one without. The one with audio sells for $7 more. I just got off the phone with a guy at Atacom (ext. 100) who swears that he has this board in stock.

[Note: this is not an endorsement of Atacom. I'm merely showing that this board is available now. Atacom's resellerrating is posted here, and it's not very good.]

The Intel i850 chipset is configured for dual-channel Rambus DRAM RIMMs. Each channel contains 2 RIMM slots, for a total of 4 per motherboard. The dual-channel configuration means that you must install your Rambus memory modules in matched pairs. So you will need at least 2 sticks of RAM. All unpopulated RIMM slots must be filled with what are called Continuity RIMMs (because Rambus is a serial technology).

Update 22jan2002 - Received a note from Dylan Williams who writes to say he has both the Asus & Abit Rambus-based boards and prefers the Asus board:

I can confirm that the P4T-E is rock solid at 140-FSB with PC800 RDRAM (1120). Benchmarks show a stock installation using 133-FSB with default voltage. The design of the P4T-E actually *helps* with RIMM overclocking. I also have a TH7II and it's nowhere as stable as the P4T-E .. altho it does have more overclocking options.

The full message is posted here. Worth reading. Despite what he says, I've heard a lot of good things about the Abit board. See here for benchmarks of related info. If you go with the Abit board, which a lot of people seem to like, make sure you get one with the ICS 9212-13 memory clock generator, which is rated for 600mhz. The older ICS 9212-03 memory clock generators are only rated for 400mhz (bad for you).

Update 04feb2002 - the Virtual Zone web site started a Northwood overclocking database (might take a while for the page to load, cuz it contains many images). Note that all the top overclockers are with the Abit board. Of course, they do not reference stability. You can download WCPUID v3.0f here. These types of images get the best quality/size ratios using GIF compression (not JPEG).


An article at Tom's Hardware, titled Pentium 4, Over 3GHz (dated 17jan2002), indicates that it might be easier to overclock DDR RAM (than RDRAM). A DDR-based Gigabyte 8IRXP motherboard was used to overclock a 2.2GHz Northwood to over 3GHz with water cooling.

These types of articles do not excite me, cuz I'm never going to install a water-cooling unit in my PC. But I was interested by a comment that I noticed on the second page:

"A motherboard with the Intel 850 chipset and Rambus memory was out of the question because RDRAM reacts very sensitively to increases in clock speed."

If that's true, that would be a significant reason to take the DDR route. I wish I could get some comparative data on relative system stability between Rambus and DDR-based systems.

Chris at GamePC seems to agree:

On our P4T-E (i850/RDRAM) motherboard, we didn't have much luck with the Northwood. On this board, we were only able to go up to 2.20 GHz, even at the motherboard's highest allowed voltage levels ... On our second platform, the Asus P4B266 (i845-D/DDR SDRAM), our results were much better. Right off the bat, we were able to clock to 2.3 GHz with only a 0.05V vCore bump. With little voltage bumps here and there, we were able to get the chip completely stable at 2.5 GHz, running on a 500 MHz FSB (125 MHz x 4).

Far as RDRAM-based systems go, I heard that you have the best chance of successful over-clocking with RIMMs manufactured by Samsung. It should be noted that the P4 was designed with RDRAM in mind .. and later adapted to DDR. That would give an advantage to RDRAM.

Many people seem to feel that most Samsung/Cosair PC800 RDRAM can hit PC1066 speeds relaibly. Anecdotal reports claim that Samsung's PC800 RDRAM max'es out at ~PC1120 to PC1160 speeds. If you need 512MB, it might be better to get 4x128, than 2x256 RDRAM sticks, as the 128MB sticks are reportedly more reliable overclockers.

While we're discussing DDR RAM, you should probably know about SiS chipsets. If stability is your #1 priority, you probably want to stick with the more proven Intel chipsets. I have not heard many bad bad reports about the SIS-based boards. They will save you a little money. Realize that time (spent troubleshooting) is also worth something.


Update 03nov2001: I've received a mailbox full of responses to my comments about Rambus RAM. Admittedly, I haven't researched RDRAM very much. All I know is that it provides performance comparable to that of DDR RAM, yet costs significantly more.

I posted one such response. This particular reader makes compelling points, and includes a brief history of RAM development. At the very end, he suggests that the (Rambus) RAM may be responsible for the P4-platform's superior stability over Athlon-based systems (which ran DDR RAM).

Since I don't mind paying extra for stability, I found this especially interesting. I asked if he had any references or links to support his claims, but have not yet heard back from him. If you'd like to see what I'm talking about, see here => Rambus RAM.


These remainder of this page contains a folksy comparison (contrast) between Rambus and DDR RAM. If you have no need for this info, skip to the next (last) page for more info related to the Intel Northwood P4 CPU.

The difference between the way that Rambus RAM & DDR RAM transfers data can be compared to the way people travel on a high-speed commuter train (on narrow rails) vs the way they travel in cars on a freeway (on a wide path, with many cars abreast).

The train (Rambus) is narrower, but moves along at a much faster rate (800MHz). Conversely, cars on the freeway (DDR) take up a much wider (data) path, but move along at a slower rate (266MHz). Both methods transfer roughly the same number of people (data) per unit time (bits/sec). Rambus RAM will transfer 3.2GB/s of data (1.6GB/s per channel times 2 channels).

3.2GB/sec is equivalent to transferring the data contained in 5 full CDs (650MB per CD) every second. That's roughly triple the bandwidth performance of my current system. Like the narrow, high-speed commuter train, Rambus RAM is only 16-bits (2-bytes) wide, but moves along at a higher clock-rate (800 MHz). DDR RAM is 64-bits (8 bytes) wide, but moves along more slowly (266MHz).

Both methods accomplish roughly the same performance, but each goes about it in a different way. DDR RAM moves more data each step (Hz), but steps along at a slower rate.

Narrow bus data transfers are known as serial. Wide bus data transfers are known as parallel. Each has its own set of pro's & con's. Parallel is more established. Serial seems to be becoming increasingly popular, and is supposedly less expensive to manufacture.

Perhaps it might be worth noting here that Rambus RAM modules, running at 800MHz, run at exactly twice the (clock) rate of the front side bus (FSB), which runs at an effective clock-rate of 400MHz. This is where the word synchronous comes in. This guide continues and concludes on the next page. Linkage below.