Will PC3200 RAM make a difference?

[dk_]

Will upgrading from a PC2700 512MB RAM stick,
to a PC3200 512 RAM stick make any noticable
difference in performance.

My computer is running with a Celeron D 340, 533 MHz FSB.

-Dennis

 

[Yousuf Khan]

Will upgrading from a PC2700 512MB RAM stick,
to a PC3200 512 RAM stick make any noticable
difference in performance.

My computer is running with a Celeron D 340, 533 MHz FSB.

None.

Yousuf Khan

 

[dk_]

None.

Yousuf Khan

Thank you for the quick reply!

Why would the faster RAM make no difference in this case?

Thanks.

-Dennis

 

[YKhan]

It's only a 15% difference in speed between PC2700 and PC3200.
Sometimes you don't even notice a difference with a doubling of speed
in some components, let alone 15%.

You'll probably be able to demonstrate benchmarks which will prove it's
faster, but you still won't notice it in your day-to-day use. There's
two different measures of performance in most computer components: (1)
bandwidth, and (2) latency. Bandwidth is akin to a topspeed, while
latency is akin to an acceleration. In a car you feel its acceleration
much more so than you feel its topspeed. Switching from PC2700 to 3200
is just an increase in topspeed, but not in acceleration.

 

[dk_]

It's only a 15% difference in speed between PC2700 and PC3200.
Sometimes you don't even notice a difference with a doubling of speed
in some components, let alone 15%.

You'll probably be able to demonstrate benchmarks which will prove it's
faster, but you still won't notice it in your day-to-day use. There's
two different measures of performance in most computer components: (1)
bandwidth, and (2) latency. Bandwidth is akin to a topspeed, while
latency is akin to an acceleration. In a car you feel its acceleration
much more so than you feel its topspeed. Switching from PC2700 to 3200
is just an increase in topspeed, but not in acceleration.

Very clear and very helpful!

Thank you.

Now here's my dilemma: ...my computer, (Celeron D 340, 533 MHz FSB),
came with one 512 MB PC2700 stick with 8 modules on it. I also have a
512 MB PC3200 stick with 16 modules on it.

I want to run the system with a total of 1024 MB's, and I understand
that I should not mix x8 with x16 sticks, so I now need to puarchase
either a 512 MB (x8) PC2700, or a 512 MB (x16) PC3200. (Currently the
PC3200 is cheaper.)

....so, does the 533 MHz FSB actually use the faster RAM (PC3200), or
does it top out with the PC2700???


Thanks.

-Dennis

 

[YKhan]

Actually, that's not a hard and fast rule about not mixing single-sided
and double-sided ram with each other. It really depends on the chipset
and BIOS you get. It's a conservative rule, but not a general rule. It
does often happen that single- and double-sided ram may cause problems,
but the only way you're going to find out if it will cause problems on
your system is to actually try it out.

You should be able to mix the old 512M DIMM with the new one, but they
will both run at the slower stick's speed (lowest common denominator).

As for the proper ram speed for a 533Mhz FSB, it actually corresponds
to PC2100 (i.e. 533 * 4 = 2132 ~ 2100); while PC3200 corresponds to
800Mhz FSB (800 * 4 = 3200). PC2700 corresponds to a 666Mhz FSB, which
Intel just completely skipped, however AMD built processors which
corresponded to that speed.

Yousuf Khan

 

[dk_]

Actually, that's not a hard and fast rule about not mixing single-sided
and double-sided ram with each other. It really depends on the chipset
and BIOS you get. It's a conservative rule, but not a general rule. It
does often happen that single- and double-sided ram may cause problems,
but the only way you're going to find out if it will cause problems on
your system is to actually try it out.

I read on the Hewlett Packard site, 'not to mix' 8x and 16x for my
computer.

You should be able to mix the old 512M DIMM with the new one, but they
will both run at the slower stick's speed (lowest common denominator).

If I understand correctly what you explained below, ...running with
PC3200 would add no value at all to my machine which has a FSB speed of
533MHz.

I understand what you wrote below to mean that no speed above the PC2100
sticks would add any value. ...Right?

As for the proper ram speed for a 533Mhz FSB, it actually corresponds
to PC2100 (i.e. 533 * 4 = 2132 ~ 2100); while PC3200 corresponds to
800Mhz FSB (800 * 4 = 3200). PC2700 corresponds to a 666Mhz FSB, which
Intel just completely skipped, however AMD built processors which
corresponded to that speed.

Yousuf Khan

Thanks once again.

-Dennis

 

[Yousuf Khan]

I read on the Hewlett Packard site, 'not to mix' 8x and 16x for my
computer.

Well, if it's been specifically written as such, then it must be thus.
Usually the suggestion not to mix single-sided and double-sided memory
are given anecdotally. For example:

Q: "Hi, I'm having trouble with my double-sided ram".

A: "Well, your chipset might not be designed for that double-sided
memory, switch to single-sided".

If I understand correctly what you explained below, ...running with
PC3200 would add no value at all to my machine which has a FSB speed of
533MHz.

I understand what you wrote below to mean that no speed above the PC2100
sticks would add any value. ...Right?

It's a little bit more complicated. The memory controller inside your
chipset is likely a "dual-channel" controller. That means that it can
take two DIMMs and make them act sort of like one faster DIMM. With a
single PC2100 DIMM, you're not maxing out your controller; you need two
PC2100 DIMMs to max it out. The two PC2100's are sort of combined into a
virtual PC4200 (real PC4200's don't actually exist). So a single PC2700
or PC3200 aren't really maxing out your memory controller's maximum
bandwidth either, but they're much closer to it than a single PC2100 is.
But when you combine a PC2700 and a PC3200, you're not going to get a
virtual PC5400 (dual PC2700) let alone a virtual PC5900 (PC2700 +
PC3200), you're stuck at a maximum of virtual PC4200 because that's as
fast as your FSB will go.

As an aside, this is really one of the main advantages to the new AMD64
designs -- they've completely eliminated the FSB, as they've built the
memory controller directly into the CPU rather than into the chipset.
The memory controller will talk to the rest of the CPU at full internal
CPU speeds instead of through the fixed FSB speed. This carries the
major advantage of not only increasing the bandwidth but also -- and
more importantly -- reducing the latency. Intel is expected to (forced
to?) adopt this design by 2007, because it will be four years behind AMD
at that point (because AMD will have had this since 2003).

Yousuf Khan

 

[dk_]

Well, if it's been specifically written as such, then it must be thus.
Usually the suggestion not to mix single-sided and double-sided memory
are given anecdotally. For example:

Q: "Hi, I'm having trouble with my double-sided ram".

A: "Well, your chipset might not be designed for that double-sided
memory, switch to single-sided".

I see.

This from the Hewlett Packard site...

"The following requirements must be met for
the DDR memory to function in Dual Channel mode:

* Same Density (128MB, 256MB, 512MB, etc.)
* Same DRAM chip technology (x8 or x16)
* All either single-sided or dual-sided
* Matched in both Channel A and Channel B memory channels"



I understand.

It's a little bit more complicated. The memory controller inside your
chipset is likely a "dual-channel" controller. That means that it can
take two DIMMs and make them act sort of like one faster DIMM. With a
single PC2100 DIMM, you're not maxing out your controller; you need two
PC2100 DIMMs to max it out. The two PC2100's are sort of combined into a
virtual PC4200 (real PC4200's don't actually exist). So a single PC2700
or PC3200 aren't really maxing out your memory controller's maximum
bandwidth either, but they're much closer to it than a single PC2100 is.
But when you combine a PC2700 and a PC3200, you're not going to get a
virtual PC5400 (dual PC2700) let alone a virtual PC5900 (PC2700 +
PC3200), you're stuck at a maximum of virtual PC4200 because that's as
fast as your FSB will go.

Very clear. Thank you, thank you.

As an aside, this is really one of the main advantages to the new AMD64
designs -- they've completely eliminated the FSB, as they've built the
memory controller directly into the CPU rather than into the chipset.
The memory controller will talk to the rest of the CPU at full internal
CPU speeds instead of through the fixed FSB speed. This carries the
major advantage of not only increasing the bandwidth but also -- and
more importantly -- reducing the latency. Intel is expected to (forced
to?) adopt this design by 2007, because it will be four years behind AMD
at that point (because AMD will have had this since 2003).

Yousuf Khan

Here's some numbers from an advertisement from e-machine's site...
"AMD Athlon 64 3200+ Processor (512KB L2 cache, 2.2GHz, 1600MHz FSB)".

The numbers show an FSB speed, (which is much faster than what is
advertised in an equivalently priced Celeron D machine.

Thanks for the help and education.

-Dennis

 

[Nate Edel]

Here's some numbers from an advertisement from e-machine's site...
"AMD Athlon 64 3200+ Processor (512KB L2 cache, 2.2GHz, 1600MHz FSB)".

The numbers show an FSB speed, (which is much faster than what is
advertised in an equivalently priced Celeron D machine.

That's not the real FSB speed, that's an estimate of the speed of the
Hypertransport bus.

Athlon 64 3200+ could be either a Socket 939 (dual channel) or Socket 754
chip. Either one has a 200mhz DDR (= "400mhz / DDC 400 / PC3200") FSB to
memory. In the case of the Socket 754, that is single channel, while the
Socket 939 chip supports dual channel memory so that it's the equivalent of
an 800mhz SDR bus (or the QDR 200mhz "800FSB" bus on recent Pentium 4
models, outside of the few 1066FSB EE's)

The raw bandwidth on Socket 939 and 800FSB P4s is similar; because of the
on-die memory controller, the latency for the AMDs is better. Then again,
even with Socket 754 (half the total bandwidth), the latency is better.

One other factor with bandwidth is that for the Athlons, there is a separate
memory bus (through the on-die controller) and I/O bus with the HT
connection to the AGP/PCI/PCI-E slots. On Intel chips, all this
bandwidth is shared between the memory and the I/O slots. In practice, I'm
not sure if this really makes a difference on single-CPU systems, although
having a separate connection for cache coherency and as many memory
controllers as there are CPUs definitely cab make a big difference for SMP
applications.

 

[YKhan]

Here's some numbers from an advertisement from e-machine's site...
"AMD Athlon 64 3200+ Processor (512KB L2 cache, 2.2GHz, 1600MHz FSB)".

The numbers show an FSB speed, (which is much faster than what is
advertised in an equivalently priced Celeron D machine.

That's a "little white marketing lie". Since most people are used to
seeing a FSB speed rating, Emachines just took the closest thing they
could find to a front-side bus, which is the Hypertransport link, and
called it the FSB. The more proper answer to what is the FSB speed of
an AMD64 processor would be "not applicable", but of course that would
generate questions like "why is it not applicable?", or "if it doesn't
have a FSB, then how does it work?", etc.

The Hypertransport link does fulfill one of the functions of the
traditional FSB, which is that it connects the i/o chipset to the CPU.
However, the memory controller is not on the i/o chipset like with
Intels, the memory controller is inside the AMD CPU, so the only job an
AMD i/o chipset does is interface with peripheral devices like video
cards, hard disks, USB, etc. The AMD memory controller is not part of
the Hypertransport connection, it is its own independent connection.
That's why AMD has coined the marketing term "Direct Connect
Archictecture" to describe the combination of its internal memory
controller, and the Hypertransport link.

Also Hypertransport is not a bus in the strict sense of the word, it is
a point-to-point link, i.e. one-to-one only. Secondly, the speed rating
given to it in the ad, "1600MHz FSB", is just a maximum possible speed;
the HT link being point-to-point is a negotiated speed. So it only
connects at the highest speed common to the CPU and the chipset it is
connecting to. If the chipset can only go as fast as 1066MHz, then
that's the speed of the HT link that they will negotiate.

Yousuf Khan

 

[George Macdonald]

Very clear and very helpful!

Thank you.

Now here's my dilemma: ...my computer, (Celeron D 340, 533 MHz FSB),
came with one 512 MB PC2700 stick with 8 modules on it. I also have a
512 MB PC3200 stick with 16 modules on it.

I want to run the system with a total of 1024 MB's, and I understand
that I should not mix x8 with x16 sticks, so I now need to puarchase
either a 512 MB (x8) PC2700, or a 512 MB (x16) PC3200. (Currently the
PC3200 is cheaper.)

...so, does the 533 MHz FSB actually use the faster RAM (PC3200), or
does it top out with the PC2700???

I'm afraid this whole discussion has gone completely astray, partly due to
lack of detailed info to start with.

First the x8 and x16 you mention in HP specs elsewhere in the thread does
not mean the number of chips on the DIMM - it's the width of the Data Out
of the memory chips used.

The HP specs you quoted are also talking about *dual* channel operation,
which it was not clear your mbrd supported to start with... but apparently
it does. Obviously with only one DIMM you are not running dual channel
with only one DIMM inserted.

For a 533MHz FSB, with a dual channel memory setup, the minimum speed match
for memory modules is PC2100 (actually 8*266.6 = 2133.3). Two channels of
those (4266.6) would match exactly your FSB of 533MHz (8*533.3 = 4266.6).
HP, however has used PC2700 DIMMs, possibly because they are the most
available.

The bottom line is that the best peformance will be obtained with the dual
channel setup and to get that you must have matched DIMMs, one (or two for
larger memory capacity according to slots) on each channel. What does
matched mean?... the same chip count, speed rating and chip arrangement on
each pair of DIMMs across the two channels. Whether you could make it work
with identically spec'd DIMMs from two different mfrs is a toss-up but I
would not even try it myself.

Your best bet: take a close look at the chips on the original DIMM and note
the mfr name and see if you can find a similar DIMM from an online vendor,
otherwise the only option is to cough up the HP price for a match. You
could post the chip codes here and soemone will likely be able to decode
them for you.

Also, check out www.crucial.com and enter the details of your system to see
what memory type and timings they recommend for it.

 

[dk_]

OK. Here's where I'm at now... I opened the new machine again and I took
out the single 512 MB stick and it is not a PC2700, but rather a PC3200,
(i.e., 400 MHZ DDR, double sided with 16 chips).

Today I purchased another 512 MB, 400 MHz DDR, double sided with 16
chips (a PC3200).

HP says on the box that it was equipped with a PC2700, but you're right
I'm sure, they actually used a PC3200 because of availability and cost.

I installed both PC3200's and booted the computer and looked in the
BIOS. The screen says there are two PC2700 sticks loaded, for a total of
1 gig total memory.

The motherboard has a FSB of 533 MHz, and according to this thread, the
motherboard will only use speed up to the PC 2100 (HP says PC2700). So
why the heck is the BIOS showing #2 PC2700 sticks, when it should
probably read #2 PC2100 sticks even though the machine actually is
loaded with #2 PC3200 sticks???!@#$%^&

BTW: where do these 4* or 8* multipliers come from?

Regarding the discussion below about x8 and x16... you point out that
these 2 numbers don't refer to the number of chips on a stick; I
haven't see any FAQ's, or discussions, using those two numbers, so I
assumed the numbers were referring to the obvious number of chips on a
stick and that they should be the same # on each stick.

The two sticks that are currently in the machine, are physically matched
(at least).

Q). How much of a difference would it make in performance if the two
chips were actually identically matched as a set, as you recommend?
(Currently by most criteria the two chips are matched, i.e., 512 MB, DDR
400, double-sided with #16 chips on each stick.)

(I did notice that in the store, and on a RAM manufacture's site that
they do talk about, and sell "dual DDR's" as a tested set.

My motherboard is an 'MSI MS-6577 version 4.1' running a Celeron D 340
if that adds any useful information.

-Dennis

 

[George Macdonald]

OK. Here's where I'm at now... I opened the new machine again and I took
out the single 512 MB stick and it is not a PC2700, but rather a PC3200,
(i.e., 400 MHZ DDR, double sided with 16 chips).

Today I purchased another 512 MB, 400 MHz DDR, double sided with 16
chips (a PC3200).

HP says on the box that it was equipped with a PC2700, but you're right
I'm sure, they actually used a PC3200 because of availability and cost.

I installed both PC3200's and booted the computer and looked in the
BIOS. The screen says there are two PC2700 sticks loaded, for a total of
1 gig total memory.

The motherboard has a FSB of 533 MHz, and according to this thread, the
motherboard will only use speed up to the PC 2100 (HP says PC2700). So
why the heck is the BIOS showing #2 PC2700 sticks, when it should
probably read #2 PC2100 sticks even though the machine actually is
loaded with #2 PC3200 sticks???!@#$%^&

I haven't studied the details of recent Intel chipsets to see how far the
FSB and memory channel can be run run out of synch; it's also possible that
the BIOS of an OEM mbrd is artificially limited like that. IOW with a FSB
of 533MHz, there's no point to PC3200 memory but since they sell the system
with only one DIMM, thus eliminating dual channel mode in the base config,
they allow the memory to run up to 2700, out of synch, to allow a little
performance boost.

These are the kinds of things you run into with OEM products like that -
artificially limited mbrds... either through BIOS limitations or through
being engineered to a price point. It's after you get fed-up with this
crap that you decide to DIY and build your own system from chosen
components, like most of the people in this NG.

BTW: where do these 4* or 8* multipliers come from?

You mean the "8*266.6 = 2133.3" I used? It's the byte width of the FSB and
memory channels, to translate MT/s to MB/s. I don't see a 4*.

Regarding the discussion below about x8 and x16... you point out that
these 2 numbers don't refer to the number of chips on a stick; I
haven't see any FAQ's, or discussions, using those two numbers, so I
assumed the numbers were referring to the obvious number of chips on a
stick and that they should be the same # on each stick.

For same-size DIMMs, the effect is the same of course. If you look up any
memory chip Data Sheets, you'll see that they commonly come in x8 and x16
data widths for desktop system DIMMs and x4 and x32 for other applications.

The two sticks that are currently in the machine, are physically matched
(at least).

Q). How much of a difference would it make in performance if the two
chips were actually identically matched as a set, as you recommend?
(Currently by most criteria the two chips are matched, i.e., 512 MB, DDR
400, double-sided with #16 chips on each stick.)

If the system is running properly and in dual channel mode, it means that
the two DIMMs are "matched enough". You might try running Prime95's
Stress Test to verify system integrity: http://www.mersenne.org/

Do you have any way to *know* that dual channel mode is actually working?

(I did notice that in the store, and on a RAM manufacture's site that
they do talk about, and sell "dual DDR's" as a tested set.

My motherboard is an 'MSI MS-6577 version 4.1' running a Celeron D 340
if that adds any useful information.

That must be a OEM-only product - it does not show at the MSI Web site, so
could even be a special HP-only product.

 

[dawg]

No. Not unless you actually raise the FSB. PC3200 just means the ram is
qualified to work UP TO 400(200) FSB. PC2700 is qualified to work UP TO
333(166) FSB. I believe your Celeron D works at a 133(133x4 for intel=533)
FSB. AMD uses a different front side bus(x2) than intel which uses a x4 bus
although they can both use the same memory.Weird huh?

 

[dannysdailys]

As you can tell by all the posts, this is the problem with buying
manufacturers PC. They're simply a mix and matched mess of parts

Assuming the MSI is using an Intel chipset: As far as I'm aware
Intel doesn't even make a dual channel memory bus. That's Nvidia
domain with the Nforce set of northbridge chips. So matching memor
seems silly to me. Maybe I'm misinformed

Even at that, Nvidia's dual channel mode really only helps for o
board video. It's less then a 3% increase without it. I found tha
out after I blew the money on a "matched" pair, put my computer i
dual channel mode, and noticed NO speed improvement

Always remember, manufacturers are selling to the lowest commo
demoninator and say all kinds of things, if it will help avoid a tec
support call

I'm sorry you threw out your old RAM chip, because it probably woul
have worked fine, even though it wasn't a direct match. You shoul
have popped it in and see if the computer would boot

In a single channel environment, you can string any RAM module
together. The only caviet being: Your RAM bus will run only as fas
as the slowest module installed

The only quirk I'm aware of doing this, is having to have the larges
module in slot one. Otherwise, the system will only recognize what'
in slot one, to two. Example: If you have a 128 in slot 1, and a 25
in slot 2; the system will only see 128 in slot 2

Maybe I'm missing something..

 

[George Macdonald]

As you can tell by all the posts, this is the problem with buying a
manufacturers PC. They're simply a mix and matched mess of parts.

Well some of them are built to spec - the lowest for the price.

Assuming the MSI is using an Intel chipset: As far as I'm aware,
Intel doesn't even make a dual channel memory bus. That's Nvidias
domain with the Nforce set of northbridge chips. So matching memory
seems silly to me. Maybe I'm misinformed.

Just a wee bit out of date. Intel has had DDR dual memory channel
chipsets for ~2years now - they need it to keep up with the 4x clocked FSB.
IIRC, as far back as the DRDRAM chipsets, they were dual channel too,
except for the i820.

 

[dk_]

Regarding the discussion below about x8 and x16... you point out that
these 2 numbers don't refer to the number of chips on a stick; I
haven't see any FAQ's, or discussions, using those two numbers, so I
assumed the numbers were referring to the obvious number of chips on a
stick and that they should be the same # on each stick.

For same-size DIMMs, the effect is the same of course. If you look up any
memory chip Data Sheets, you'll see that they commonly come in x8 and x16
data widths for desktop system DIMMs and x4 and x32 for other applications.[/QUOTE]

Thank you for all the info and details.

I have been searching to find how I can identify or recognize 'data
widths' for each DRAM stick. How can I tell if the stick is x8 or x16.

In order for dual channel mode to work, some sources say the 'data
widths' must match, and some sources make no mention of this.

I have searched the term 'DRAM bus width', and so far I can find no
practical informaltion.

Help.

Thanks again.

-Dennis

 

[Jason Gurtz]

Just a wee bit out of date. Intel has had DDR dual memory channel
chipsets for ~2years now - they need it to keep up with the 4x clocked FSB.
IIRC, as far back as the DRDRAM chipsets, they were dual channel too,
except for the i820.

Other than server chipsets (hmm, maybe that was all serverworks anyway),
wasn't the 865 the first from Intel to support dual channel?

~Jason

--

 

[George Macdonald]

Other than server chipsets (hmm, maybe that was all serverworks anyway),
wasn't the 865 the first from Intel to support dual channel?

Well, in the strictly "desktop" arena that's probably true... if you figure
the i840 & i850 were at least targeted at a higher level. OTOH neither
ended up in many servers and they migrated down from even the workstation
slot to what I'd call the high-end desktop space. You gotta figure there
were a lot of people who did not want the full compromise of i810/i815 and
even i845 - Intel dropped the ball there... and DRDRAM *was* intriguing to
some. They just had to feel it for themselves.