Sign in with
Sign up | Sign in
Your question

2 x 256 or 1x 512

Last response: in Memory
Share
Anonymous
a b } Memory
July 27, 2001 6:15:00 AM

Do you all reccomend getting 2 x 256 MB PC2100 RAM or 1 x 512 MB PC2100 RAM?

More about : 256 512

July 27, 2001 6:55:39 AM

That would depend on what kind of future upgradability you'd like and how many slots there are available. You most likely have 3 slots for memory. If you go with the two 256MB modules, you will only have one free slot for expansion. If it were me, I'd go with the 512MB module. (Actually if it were me, I'd max out the memory on the motherboard immediately. :) 

-Raystonn

= The views stated herein are my personal views, and not necessarily the views of my employer. =
July 27, 2001 3:23:52 PM

True, but if you get two sticks and then get the nForce, you can take advantage of the Dual-channel RAM.

But since RAM is so cheap now, I'd say go for one 512 (even though it's more expensive than two 256s)

-----------------
Whoever thinks up a good sig for me gets a prize :wink:
Related resources
July 27, 2001 7:55:30 PM

If you only have one PC and you life depends on it, get 2 sticks of 256. This way if one of the sticks die on you, you can still limp on with the other stick of 256. If you doubt this, ask me. Damn that Murphy's Law.

It don't come easy.......not always.
July 27, 2001 8:17:06 PM

A very good point.

-----------------
Whoever thinks up a good sig for me gets a prize :wink:
July 27, 2001 8:42:41 PM

The nForce chipset is only dual-channel in that one channel (and hence one memory slot) is dedicated to the integrated graphics chipset while the rest is dedicated to the CPU and the rest of the system. The memory you place in the slot that's dedicated to the graphics chipset will be unusable as standard system memory by the CPU. In essence, they removed all memory from the integrated graphics chipset and instead allow you to determine how much onboard video memory you'd like yourself, which you can plug into the graphics memory slot. The actual memory usable by the real CPU is still using a single channel. You will in fact see absolutely no benefit with this extra graphics memory channel if you choose to install an add-on video card, as most performance-minded people do. It will sit there being entirely unused.

-Raystonn

= The views stated herein are my personal views, and not necessarily the views of my employer. =
July 27, 2001 9:13:36 PM

Really? I've never heard that. I'll have to go back to actually learning about stuff before I post :) 

-----------------
Whoever thinks up a good sig for me gets a prize :wink:
July 27, 2001 9:21:16 PM

Quote:
NVIDIA has given the results of the nForce in the Quake3 (1024X68X32) in the presentation. The nForce's performance with 64bit memory access is nearly the same as that of GeForce2 MX200 based cards, and with the 128bit memory access the nForce performs nearly 25% faster than the GeForce2 MX200.


That's just bad.
Good source for info on the nForce <A HREF="http://www.digit-life.com/articles/nvidianforce/" target="_new">here</A>

-----------------
Whoever thinks up a good sig for me gets a prize :wink:
July 27, 2001 9:27:28 PM

Quote:
you can still limp on with the other stick of 256

True, but I'm a bit worried about any system which <i>limps</i> with 256MB of RAM. Very high end systems accepted, this amount of RAM is still going to handle things nicely.

<b>
"Now drop your weapons or I'll kill him with this deadly jelly baby." :wink:
</b>
July 27, 2001 10:00:39 PM

He was just making a point, I don't think he meant the word "limp" seriously.

Or we can take it a bit farther...how can a computer limp when it doesn't have legs? :) 

-----------------
Whoever thinks up a good sig for me gets a prize :wink:
July 28, 2001 1:48:04 AM

512 is a fair bit more expensive... if you can afford it without burning a hole in your pocket go for 512 otherwise stick to the 256 sticks...

you do not strengthen the weak by weakening the strong
July 28, 2001 3:53:13 AM

My mobo has 1.5 GB RAM Max (FIC AD11). Is there any way to increase this?

When I rule the world, Apple will only mean the fruit.
July 30, 2001 8:41:59 PM

Nope, unless you want to do some heavy MB modding. (Let me give you a hint: no, you don't :) 

-----------------
Whoever thinks up a good sig for me gets a prize :wink:
July 30, 2001 10:55:42 PM

I'm afraid your information is inacurate. nForce 128 is a dual channel chipset, and the memory bandwith is shared between the system and the graphic subsystem without access lock-out, through a cross controller. As I perceive it, it has some similarities with SGI systems in that aspect.
If you information was correct then the 64 bit version of the chipset would have no memory bandwith either to the GPU or the rest of the system (or else it would use an "exquisite" 32 bit bus to both ;) )

But you still can't buy nForce based mobos... yet.


How terrible is wisdom when it brings no profit to the wise
July 31, 2001 1:11:23 AM

Sorry, I hate to be the bearer of bad news, but that's simply not the case. The dual memory channels are split between the integrated graphics chipset and the rest of the system. When you use an add-on graphics board, you do not need to put any memory into the graphics memory slot. (It would not be used.) Take a look at the nForce architecture diagram. You will see the memory split into two different locations. (One dedicated to the graphics chipset.)

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
July 31, 2001 2:42:03 PM

Don't worry, we are not in ancient egypt :wink: .

But AFAIK, there is no such thing as a graphic memory slot and there is no splitagge between the channels - they are each connected to an independent memory controler, as can be seen in these diagrams http://www6.tomshardware.com/mainboard/01q2/0106041/nfo...
http://www.anandtech.com/chipsets/showdoc.html?i=1484&p...
(this one is explicit regarding this).

Or, as Nvidia puts it (in the middle of all that marketing crap), "TwinBank's dual-independent, cross-bar memory controller allows the CPU, GPU, and APU simultaneous access to the system's 4.2GB/sec. of memory bandwidth, guaranteeing continuous access for all applications, all the time. (http://www.nvidia.com/view.asp?IO=twinbank)"
This means that the dual channel version (nForce 440?), when paired with an add-on graphics board will have 128 bit DDR access - 4.2GB/s of bandwith. When using the internal GPU, you still have the 4.2GB/s of memory bandwith, but the graphic susbsytem will consume a much larger part of it, since in this case it doesn't have its own memory bank.

Where did you get any info regarding a graphics memory slot?


How terrible is wisdom when it brings no profit to the wise
July 31, 2001 3:21:17 PM

He's talking about a video card, not add-on video memory.

-----------------
Whoever thinks up a good sig for me gets a prize :wink:
July 31, 2001 7:15:42 PM

Raystonn said:
Quote:
When you use an add-on graphics board, you do not need to put any memory into the graphics memory slot. (It would not be used.) Take a look at the nForce architecture diagram. You will see the memory split into two different locations. (One dedicated to the graphics chipset.)

1) - there is no dedicated memory location to the graphics chipset.

2) - taking 1) into account it is quite obvious that that there is no "graphics memory slot"

My interpretation - and it can obviously be wrong, is that Raystonn was saying that one of the nForce memory channels was reserved to the graphic core, and if one used an add-on adapter, that channel would go unused.
The issue is that none of the nForce channels is reserved to the on-chipset GPU - there is 4.2 GB/s of memory bandwith available for the entire system - including CPU, southbridge...

Or, in other words, in nForce there is no such thing as a free lunch,... :wink: I mean a graphic memory slot.


How terrible is wisdom when it brings no profit to the wise
August 1, 2001 2:03:45 AM

I don't think I'm being very clear so I'll take another approach. What I'm attempting to point out is that the CPU does not have access to all of this bandwidth. Half of the bandwidth (in essence, a full channel's worth of bandwidth) is indeed meant for the graphics chipset. Looking past nVidia's marketing for a moment we can clearly see this in their technical brief. Anandtech's picture of the layout (from your Anandtech url) was ripped (without credits I might add) from nVidia's technical brief pdf file, which can be found <A HREF="http://www.nvidia.com/docs/lo/49/SUPP/nForce_TwinBank_M..." target="_new">here</A>.

If you'll take a magnifying glass to the one on Anandtech's site (or look at the PDF on page 5 of 10) you'll notice they give maximum bandwidth figures between the various components. Take a look at the figure for the bus that connects the CPU and the memory system. It offers a "whopping" 2.128GB/sec of memory bandwidth for your CPU. That figure happens to match the bandwidth of a single memory channel exactly. What's important here is the amount of memory bandwidth offered to your CPU. We can clearly see from the same picture that an AGP video card will only consume 1.064GB/sec of memory bandwidth at most. This is pretty much the same for all systems currently in existence with a 4x AGP slot. You now have another 1.064GB/sec of memory bandwidth left over that cannot be utilized by the CPU. This is wasted memory bandwidth. Your CPU will still get only 2.1GB/sec, the same as currently offered by any other PC2100 DDR motherboard. Is this what their promotional material says? It doesn't look that way to me. Does it provide more bandwidth to the CPU than the current dual channel RDRAM designs (3.2GB/sec)? No.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
August 1, 2001 3:25:53 AM

This from the <A HREF="http://www.amd.com/products/cpg/athlon/keyfeatures.html" target="_new">AMD Web Site:</A>

"... 266MHz or 200MHz AMD Athlon™ processor system bus enables leading-edge system bandwidth for data movement-intensive applications ... Peak data rate of 1.6 to 2.1GB/s (depending on processor bus speed)..."

It appears that the Athlons have a theoretical max bandwidth of 2.1GB/s.

This from the <A HREF="http://developer.intel.com/design/Pentium4/prodbref/#bu..." target="_new">Intel Web Site:</A>

"...
400 MHz System Bus
The Pentium® 4 processor supports Intel's highest performance desktop system bus by delivering 3.2 GB of data per second into and out of the processor. This is accomplished through a physical signaling scheme of quad pumping the data transfers over a 100-MHz clocked system bus and a buffering scheme allowing for sustained 400-MHz data transfers. This compares to 1.06 GB/s delivered on the Pentium® III processor's 133-MHz system bus.
..."

It appears that the P4s have a theoretical max bandwidth of 3.2GB/s.

I still think dual DDR banks would be nice. So the nForce board still represent an attractive option for me, performance wise. My assumption is that there will be no CPU memory access slowdowns when the graphic card is accessing memory.

Thanks for shedding light on the subject.
August 1, 2001 3:38:49 AM

"It appears that the Athlons have a theoretical max bandwidth of 2.1GB/s"

The CPUs do not have a maximum bandwidth. Buses do. Current memory architectures offer 2.1GB/s for PC2100 DDR.


"My assumption is that there will be no CPU memory access slowdowns when the graphic card is accessing memory."

Yes. This will be a first for an integrated graphics chipset. However, all add-on cards have their own memory and have never incurred penalties when accessing their memory. Thus, you gain nothing when using an add-on video card. It's the same as all the rest of the chipsets/motherboards; just another 2.1GB/s PC2100 DDR motherboard.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
August 1, 2001 4:08:43 AM

I mean the bus that the processor uses is limiting the bandwidth that it can utilize. Giving the system more total bandwidth won't enable the processor to use it, hence the 2.1GB/s CPU bandwidth limitation from/to the CPU in Figure 3 of the "nForce TwinBank Memory Architecture Tech Brief" document that you referred us to.
August 1, 2001 4:10:44 AM

That bus is actually on the motherboard, not the CPU. It's part of the nForce chipset.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
August 1, 2001 4:51:33 AM

There are specified limits to how fast data can flow in or out of the Athlon processor. According to AMD documentation, that specified limit is 2.1GB/s (for 266FSB). According to AMD, the Athlon should only interface with a 2.1GB/s max bus. Total memory bandwidth for the system is still 4.2GB/s, only 2.1GB/s is visible to the CPU. The fact that the CPU only sees a 2.1GB/s bus does not imply that only one bank is dedicated to the CPU which you seem to suggests:

"...The nForce chipset is only dual-channel in that one channel (and hence one memory slot) is dedicated to the integrated graphics chipset while the rest is dedicated to the CPU and the rest of the system. The memory you place in the slot that's dedicated to the graphics chipset will be unusable as standard system memory by the CPU..."

and

"...Take a look at the figure for the bus that connects the CPU and the memory system. It offers a "whopping" 2.128GB/sec of memory bandwidth for your CPU. That figure happens to match the bandwidth of a single memory channel exactly. What's important here is the amount of memory bandwidth offered to your CPU. ..."

and

"...Take a look at the nForce architecture diagram. You will see the memory split into two different locations. (One dedicated to the graphics chipset.)..."
August 1, 2001 5:00:57 AM

"There are specified limits to how fast data can flow in or out of the Athlon processor"

Actually, there aren't. It all depends on the speed of the bus to the CPU. Obviously a CPU is limited by the clockspeed at which it is rated. The clockspeed is determined by the bus frequency multiplied by the CPU's "multiplier". If you decrease the CPU multiplier and increase the bus speed a proportional amount, you have just increased the amount of data that can be sent to the CPU without affecting the actual speed of the CPU itself (its clockspeed.)


"The fact that the CPU only sees a 2.1GB/s bus does not imply that only one bank is dedicated to the CPU which you seem to suggests"

Both memory channels are in fact available all components, including the CPU. I apologize for stating otherwise. I had seen another diagram a few months ago that had led me to this conclusion. However, that doesn't change the fact that you have the bandwidth equivolent of only a single channel available to the CPU. That is 0 progress in terms of memory bandwidth.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
August 1, 2001 7:04:02 AM

Quote:
It offers a "whopping" 2.128GB/sec of memory bandwidth for your CPU

This is a limitation of the current implementation of Athlon EV6 based bus protocol (133Mhz DDR), not the chipset, nor is it connected to a memory channel. If each channel was able to provide 2.7GB/s the 2.128GB/s would remain as long the system bus isn't changed. Also the 1.064GB/sec is limitation of the AGP bus, valid for all chipsets.
Quote:
You now have another 1.064GB/sec of memory bandwidth left over that cannot be utilized by the CPU. This is wasted memory bandwidth.

Not quite - It would be all wasted if it wasn't for a "small" detail :wink: - the southbridge - all PCI devices and on chip multimedia (not that they are able to use all the available bandwith)...

Current PC2100 DDR aren't able to garantee to the CPU all bandwith it can use - as the 2.128GB/sec of memory bandwith it can require is always shared with the other system devices - namely AGP, than can consume up to half of it. By providing twice the system bandwith, the CPU, in practice, doesn't have to compete for bandwith - it is always available. In its present version is there some bandwith that goes unused (with add-on graphics) - yes there is (something between 264MB/s and roughly 931MB/s).

Quote:
Does it provide more bandwidth to the CPU than the current dual channel RDRAM designs (3.2GB/sec)? No.

Likewise I could sustain that the dual channel RDRAM never provides the cpu with 3.2GB/s since some of this bandwith is used by AGP and other system devices (up to roughly 1.3 GB/s), never being completely available for the cpu.


And BTW, Anadtech did rip the image without credits... naughty boy.

How terrible is wisdom when it brings no profit to the wise
August 1, 2001 7:18:20 AM

"However, all add-on cards have their own memory and have never incurred penalties when accessing their memory. Thus, you gain nothing when using an add-on video card. It's the same as all the rest of the chipsets/motherboards; just another 2.1GB/s PC2100 DDR motherboard."

Before reaching the add-on card memory banks, the data must travel from main memory through the AGP bus. You aren't sugesting that AGP memory access is "for free", and doesn't use some bandwith that otherwise could be used by the CPU...


How terrible is wisdom when it brings no profit to the wise
August 1, 2001 7:59:31 AM

Video cards have their own memory on-board. The AGP bus is only used to upload new textures and vertices to the video card's memory between levels in most games. 99.9% of the time only the memory on the actual video card is being used each frame. Most game developers know that sending new textures and vertices to the video card is slow and keep this to a minimum. This is what locked vertex arrays/buffers is all about. (Locked means it's locked into the video card's local memory, not requiring any AGP transfers to read/write that memory.)

In essence, AGP bandwidth use is negligable on today's high end video cards. But even with the AGP bus going full stream (which just doesn't happen except while loading up texture memory and such before the game actually starts), we still have about 2.2GB/s of memory bandwidth available to the CPU. 99% of the time we'll have nearly the full 3.2GB/s available. No matter how you look at it, nForce still provides less bandwidth with only 2.1GB/s for the processor.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
August 1, 2001 8:05:53 AM

"This is a limitation of the current implementation of Athlon EV6 based bus protocol (133Mhz DDR), not the chipset, nor is it connected to a memory channel. If each channel was able to provide 2.7GB/s the 2.128GB/s would remain as long the system bus isn't changed."

That's the point. They could up the bus speed between the processor and the memory controller (nForce). If you double the bus speed and halve the CPU multiplier, the CPU runs within spec and you get twice as much memory bandwidth. This is actually what most people think they would be getting from nForce. They envision having more memory bandwidth than the Pentium 4 with RDRAM but that simply isn't the case.


"And BTW, Anadtech did rip the image without credits... naughty boy."

I also found the same image used without credit on Tom's site in his review of nForce. He should be spanked as well.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
August 1, 2001 9:40:07 AM

Now we're talking :wink: . People will be getting increased bandwith but definitely not twice the bandwith.
The issue is not so much the (seemingly brilliant but lets wait for actual hardware before casting definitive judgement) nForce chipset bandwith vs i8xx bandwith but more of Athlon (and EV6) vs P4 (and whatever is the name of its bus) memory bandwith.

BTW, although the the FSB is in the motherboard it isn't in the chipset - it connects the chipset (the memory controler, or, more precisely, the northbridge) to the socket (or slot) with the CPU. Obviously, in order to change the bus you have to have the CPU maker support (usually it is even their initiative), namely due to locked multipliers. A protocol change (not frequency) usually leads to a new processor series (even relatively minor changes - like GLT+ became AGLT+). No chipset maker alone can overcome this.



How terrible is wisdom when it brings no profit to the wise
August 1, 2001 3:56:28 PM

Quote:
We can clearly see from the same picture that an AGP video card will only consume 1.064GB/sec of memory bandwidth at most. This is pretty much the same for all systems currently in existence with a 4x AGP slot.


Right. nVidia's website says that the on-board graphics have "8x architecture", or something like that, can't remember the exact words. Anyhow, it's supposed to operate like 8x, which still probably isn't as good as memory on a card, like we have now.


Quote:
I also found the same image used without credit on Tom's site in his review of nForce. He should be spanked as well.


Don't get all kinky on us, Raystonn.

-----------------
Whoever thinks up a good sig for me gets a prize :wink:
August 1, 2001 11:51:46 PM

Quote:
which still probably isn't as good as memory on a card, like we have now.

You're right in all but one esoteric case - a game with such a huge amount of textures (and "unique" programing) that would always be retrieving data from memory. In this non-existent case the "AGP8x" would give some superiority (bandwith wise) over an add-on solution (limited to AGP4x).
The bandwith of the 230Mhz DDR of top add-on boards is way over the system memory bandwith (128 x 2 x 230 / 8 = 7.36 GB/s) much less the "AGP8x".


How terrible is wisdom when it brings no profit to the wise
August 3, 2001 1:57:47 AM

hey rayston.. another quozzie here.

say i had a standard ddr or sdr system and i benchmarked:
4x128mb
2x256mb
1x512mb
assuming that all the ram has identical latencies & from the same company...
woould i ever see any performance difference? (ignoring the upcomming nforce )

regards, MasterPoo + Hamster

"i love the smell of Overclocking in the morning!" Says my Hamster.
August 3, 2001 4:49:10 AM

No. Performance would be identical.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
August 3, 2001 4:30:23 PM

Good point, arbee. There aren't many games that need more than my 64meg, but I'm sure they exist.

-----------------
Whoever thinks up a good sig for me gets a prize :wink:
August 4, 2001 9:45:56 PM

Trouble is many new boards based on the AMD 761 Chipset only have two slots for memory. So if you go with 2 x 256 you will have to get rid of memory to expand.
August 6, 2001 5:18:48 PM

Quote:
The nForce chipset is only dual-channel in that one channel (and hence one memory slot) is dedicated to the integrated graphics chipset while the rest is dedicated to the CPU and the rest of the system. The memory you place in the slot that's dedicated to the graphics chipset will be unusable as standard system memory by the CPU. In essence, they removed all memory from the integrated graphics chipset and instead allow you to determine how much onboard video memory you'd like yourself, which you can plug into the graphics memory slot. The actual memory usable by the real CPU is still using a single channel. You will in fact see absolutely no benefit with this extra graphics memory channel if you choose to install an add-on video card, as most performance-minded people do. It will sit there being entirely unused.

What a bunch of hogwash!

Fact, the athlon's bus from the cpu to memory controller has 2.1g bandwith. This fact you have gotten right, and I posted this several months ago. But from there you spew forth several things which are incorrect.

The second memory controller is not merely dedicated to graphics as you suggest but is accesable by all components. The memory to the cpu is not limited to single channel. Both banks of memory can be addresed by the cpu via the memory controller. The crossbar memory controller address's most importantly one thing and that is latency, a subject in which you completely neglect ( albeit probably not very important to you seeing how you love RDRAM). Yo need to look at the diagram in which you posted again and you can clearly see that the memory controller has access to both banks of memory. As for the actual benefit in bandwith it is true that the nforce is not true dual channel in nature as the ev6 bus as it is utilized in the athlon is only single channel as opposed to the alpha implementation that utilizes a 128 bit cpu to memory controller pathway. But to then draw the conclusion that that nforce will offer no more of performance increase than traditional motherboards is a far strectch to say the least. A traditional motherboard with a single channel memory controller would have to share its 2.1 gigs of bandwith with other devices, while the nforce offers up 4.2 gigs of bandwith to all devices insuring that that cpu to memory controller will always have its optimum bandwith and not be competeing with other devices.


A little bit of knowledge is a dangerous thing!
August 6, 2001 11:25:14 PM

Raystonn has recognized that he was misinformed.

Anyway he (albeit indirectly) raised an interesting question - the Crush chipset provides more bandwith than the system is able to use - I think the Athlon FSB could use an upgrade in frequency with the new Athlon 4 - to 166Mhz DDR.


How terrible is wisdom when it brings no profit to the wise
August 7, 2001 8:19:01 AM

Please read before you write. We've already been through this. While the memory channels are not dedicated, the CPU is still limited to 2.1GB/sec of bandwidth. The bandwidth available on a Pentium 4 system with RDRAM (3.2GB/sec) will always surpass this, whether it's being shared with the AGP slot (using about 1 GB/sec) or not. Usually most of the AGP slot's bandwidth remains unused except between levels in a game, when new textures and vertices are sent to the video card's memory. Therefore you would usually have about 3 GB/sec of bandwidth available. I fail to see what's so impressive about nForce's 2.1GB/sec.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
August 7, 2001 8:20:50 AM

Quote:
Anyway he (albeit indirectly) raised an interesting question - the Crush chipset provides more bandwith than the system is able to use

Yes and no, not when you are using the integrated graphics. To much emphasis has been placed on the bandwith of the nforce ala p4 bandwith envy. This is not the only thing the nforce brings to the table. The crossbar memory controller being the most important, then perhaps the prefetch which should bring near a4 performance to existing tbirds, then we also have the hypertransport bus as well. To make a broad sweeping statement that says it offers no real advantage over current motherboards is way of the mark. While bandwith is important, latency is as well and the nforce address's this issue.

A little bit of knowledge is a dangerous thing!
August 7, 2001 8:44:13 AM

Actually I was brought to this post of yours by a link you made in another thread, isn't it you that should have at least looked at the link you yourself posted before making an assumption? Granted, I should have probably read the entire thread in which you finally admittd you were wrong I read your post and replied to it immediatly, my apologies. Yet you seem to think that because of the bandwith from the cpu to memory controller of the nforce being limited to 2.1 gigs this makes it an ordinary board, nothing can be further from the truth. Yeah yeah yeah, bandwith of the p4 is greater look here to a thread I posted three months ago.....

<A HREF="http://forumz.tomshardware.com/modules.php?name=Forums&..." target="_new"> here</A>

Remember the video card is not the only device that access's system memory, you still have the sound card and network card and so on. The p4 has higher bandwith yes and a higher latency as well. The two work hand in hand. Your old argument of Rdram having less latency than ddr Ram ( which I never bought) when at a higher load no longers applies to this chipset. Face it Rdram is dyeing a slow painfull death, when intel finally allows motherboard makers to use full speed DDR it will all but disappear. Bye the way, didn't we have a certain bet about the cost of DDR versus the cost of RDram...Hmmmm? And I am still dyeing to know as I have yet to see any conclusive evidence, with the northwood regain its the addtional FPU that was omitted on the williamette as you previously stated or will it just have a larger cache?

A little bit of knowledge is a dangerous thing!
August 7, 2001 9:11:43 AM

With DDR 333 coming out soon you will probably see just that in the near future as long as there is a northbridge that can support that FSB speed. Now something all together different to ponder....Take a close look at the block diagram for the nforce..... Imagine replacing the Gpu ( modifying it actually) to another cpu interface, whala, one kick ass dual board and the only expense is losing the integrated graphics.

A little bit of knowledge is a dangerous thing!
Anonymous
a b } Memory
August 7, 2001 3:20:22 PM

Hi,

I agree - max out the memory as soon as you can :-)
However, I must just add a couple of things you said about the nForce. You say that most games only upload textures & geometry at each level load. This is true, only for textures. If you're talking about TnL graphics cards, all of the current ones use the AGP to stream the vertex data in, and use the card's videomemory for texture bandwidth. It gives them around 1Gig/sec dedicated vertex bandwidth. So with a separate graphics card, the nForce provides much better bandwidth to the CPU than a standard DDR motherboard.

Also, (and here Im guessing, not speaking from experience), there is the issue of the hardware prefetcher - not on the CPU (in the case of the Athlon 4), but on the motherboard. If this is situated after the northbridge, then it will have all the bandwidth of memory (well, whatever is left after AGP ;-)) to do prefetching... probably some of this is wasted, but it does still allow the maximum bandwidth to go via the northbridge.
Lastly, there is of course the issue of latency. I guess we'll have to wait until actual boards are benchmarked, but having two separate channels does lead to lower latencies for data to get to the cpu (again, a guess, not experience!)


- Davva.
August 7, 2001 10:23:59 PM

Obviously so.

BTW I've always stated that nForce, at least on paper, will bring some serious improvements despite (IMHO) being a midstep product (a *bit* like Merced vs Mckinley) - specially the northbridge. Even if we ignored the prefetch and the high speed north/south bridge link, the technical feat of developing a dual channel DDRAM bus is quite impressive.

Even if, for the time being, and when using an add-on video card :wink: , the system isn't able to fully use the extra bandwith, it is way better than the current status of being underfed by existing motherboards.



How terrible is wisdom when it brings no profit to the wise
August 7, 2001 10:57:02 PM

That is one of the possible evolutions/versions of nForce - and it would be an interesting thing, indeed.

The problem with the announced nForce is that is too expensive for the value buyer and has a useless integrated core for the performance buyer (as I'm, at this time, neither of those it would actually fit my requirements rather well :smile: ).

The main issue in implementing the dual CPU solution (besides core logic for SMP operation), may very well be motherboard traces - after all, the dual memory channels already take some mobo realestate - and a dual bus would add some more demand for space. But it would be a great box :smile: .

But focusing on the moment, there is memory oficially supporting DDR333 (PC2700), and the EV6 Bus is quite able to operate at this frequency. If the northbridge makes it...

How terrible is wisdom when it brings no profit to the wise
August 8, 2001 3:58:15 AM

"...However, that doesn't change the fact that you have the bandwidth equivolent of only a single channel available to the CPU. That is 0 progress in terms of memory bandwidth..."

Intel designed the i840 chipset for the P3 (See the video <A HREF="http://developer.intel.com/design/chipsets/840/animate...." target="_new">here</A>). Looks like dual channel RDRAM implementation providing 3.2GB/s total system bandwidth. What good is this system when the CPUs (looks like there share the same bus) can only obtain at most 1.0GB/s memory bandwidth?

Concept looks similar to that of the nForce design (i.e., the idea of providing more system bandwidth than the CPUs can possibly use).
August 8, 2001 7:00:16 AM

Absolutely no good. I will not defend a poor product on the basis of who made it. The i840 has quite a bit of unused memory bandwidth. The CPU will never benefit from this.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
Anonymous
a b } Memory
August 8, 2001 8:44:57 AM

Hi,

Yesterday I posted in reply to the original question, which I guess most people have now skipped over, so I'll pop in here :-)

There are a couple of points which I don't think are correct in respect to the nForce (and also possibly the i840)
You say that most games only upload textures & geometry at each level load. This is true, only for textures. If you're talking about TnL graphics cards, all of the current ones use the AGP to stream the vertex data in -each frame-, and use the card's videomemory for texture bandwidth. It gives them around 1Gig/sec dedicated vertex bandwidth. So with a separate graphics card, the nForce provides much better bandwidth to the CPU than a standard DDR motherboard (same is true of the i840 I guess).

However, the second point is that the nForce also does prefetching. Now, Im guessing here, and I could be entirely wrong, but isn't it likely that the prefetcher will access a fair bit more memory than the CPU (due to wasted fetches) ? If thats the case, and that the prefetcher has access to the full 4.2G bandwidth, rather than the CPU's 2.1G/sec bandwidth, then that will prove beneficial too.

Obviously the i840 is not an efficient or balanced design, as there is no way the full bandwidth can be used - but on a full load system, the nForce seems to be extremely well balanced:
2.1Gig/sec CPU
1.0Gig/sec AGP
Prefetcher ??Gig/sec
PCI activity: 133Meg/Sec

Seems pretty good to me... maybe a *little* overkill, but not by much!

- Davva.
August 8, 2001 8:19:56 PM

Actually most games these days make use of compiled vertex buffers/arrays. These are locked into the memory on your video card. You do stream it over AGP when you first lock it into place (usually between levels) but thereafter the same vertices are reused from the video card's memory.

Anyhow, the point is that the CPU still has access to only 2.1GB/sec of memory bandwidth. nForce will get you closer to that maximum 2.1 than other DDR motherboards, but still not beyond it.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
August 8, 2001 8:25:40 PM

Does anyone know about Palamino having greater bandwidth or not?

-----------------------
Quarter pounder inside
!