Mainboards with memory soldered on

[William Gaatjes]

Hello everybody.

I wonder how much difference in speed and especially latency there would exist if we could buy our motherboards with main memory soldered on to it instead of plugging it in.

I noticed that for me and most people i know , that when a new board is bought, new memory is bought too.
And that upgrading of memory alone is non existent.

I would not have a problem buying a new board with memory soldered on to it and without the possibility of upgrading memory.
If this would give me a usefull latency reduction in memory timings.
I know that subjects like termination and reflection on signal lines would be less of a problem since you have in a sense a point to point connection.
And less trace lenght too helping the variables mentioned in the previous line.

For those that like to upgrade memory there would always be the more traditional motherboard pcb with memory sockets. Everybody happy.

We are now on the transitition that both Intel and Amd are having on board memory controllers.
My view is that this al together would be a great possibility for real fast memory. Make that cache less usefull and use it for dedicated calculation engines with their own local very wide ram.

Most of the time when a new and improved processor model comes out it needs new memory technology.
And the past has shown that usually there is also a new accompanying
chipset that makes use of the speed of the processor so most of the time people would be more willling to buy the new board.

If i was able to buy a motherboard with 2GB memory soldered onto it allowing for tight but stable timings at normal low voltages then i would.
And i mean the kind of timings which i normally could only get with overclocked ram modules causing more heat



For those who start about more memory the better :
My opinion is :
When the 64 bit area starts it is truly the time for more memory then 4GB.
32 bit with memory banking tricks causes only a lot of overhead and at the event horizon this technique will start to slow down to a point of no return.
Then 64 bit linear address space(although maybe not 100% physically implemented 38 bits would be enough for a while = 256 GB ) is more of a blessing.
And the extra calculation power of 64 bit wide can be really handy too.





As i see this trend is that most people only so often upgrade their video card or buy pheriperals but when they buy a new motherboard they buy new memory too.

Anybody any thoughts about it ? I know the memory module manufacturers(kingston geil,corsair,ozc... To name a few...) would not be happy with my suggestion but i bet a lot of customers would be.



I know i would if it makes a difference. More speed at the same power envelope. And cooling maybe easier too. Just mount those ram chips on the backside of the motherboard pcb.
And give those chips some good cooling.

 

[Lorne]

Your idea is plausable but not practical, It was done in the past durring the 286/386 days and it didnt hold with technology in a constant state of change.

Though you can buy MB's today like that but they are mainly used in office systems and some video systems, But cant make it in the videogame or image rendering world, I think Via still makes a few.

But to use or try something like that in a human FRAPS world would be grosely overpriced/costly for the simplest of OC'ing, One IC gets smoked and spend a lot of money for whole new system, Not practical.

We also live in a world where wall power still isnt the cleanest without buying some line cleaner and UPS system, Ive spent my share of money on replacing ramsticks and videocards because of static/brownouts/blackouts, I hate to think what it would have cost to replace it all everytime.

As for 64bit era, Its been here for along time already, Its the programmers that have been biggest of hold backs in that area but there also hasnt been any needs to push any harder other then the need for drivers to support it.
As for bandwidth your behind the times, Most systems now support 128bit dualchannel and there are systems and GFX cards that have 256bit and more than 512bit wide.

 

[William Gaatjes]

Hello,

"Though you can buy MB's today like that but they are mainly used in office systems and some video systems, But cant make it in the videogame or image rendering world, I think Via still makes a few."

I doubt that cause when i look inside any older or current console...
For the image rendering world motherboards with memory upgrades can always still be made. And for the enthusiasts afcourse.

And most people do not overclock. For those people it would not make a difference.
I am not really overclocking either. Tried it but felt the gain was not enough for the higher heat output.

I do like the idea of i think it was Intel of where the cores allow it , they can overclock automatically but to a save maximum frequency.
I am not sure but i thought some videocards had that possibility too.
That way it is a design feature. And since the manufacturer knows best how far to push it's hardware...

But i fully understand that mainboard with separate memory should still exist. Cause as enthusiast you want one and i agree they should still be sold.

I have not ever had such a situation but if your grid is that unreliable i advice you to get an ups. I would not feel save with any kind of hardware knowing it could turn to toast in any second. Maybe your PSU is also not that thrustworthy.


The programmers in a sense yes. The Main OS manufacturer could have used 64 bit as a trumph card to sell new systems or software with new possibilities.
Unfortunately that has not happened. It is for a programmer no use to write 64 bit software for a 32 bit market. You see, you need a big enough market. And at this time a company like Microsoft has these possibilities.
64 bits is not being pushed. Apple does the most 64 bit pushing in the more common world.
windows, osx and linux are all available in 64 bit versions. But only 1 os has a rather large desktop market. 64 bit push would have been ideal to sell new hardware and loose the limitations of 32 bit in a few years. Too bad. Another missed shot.

i am aware of the current memory technologies. I should have mentioned Amd meaning the K8 and K10 cores and upcoming cores. I should have mentioned nehalem from Intel.
Since you know about bandwith you know what kind of bandwiths they will be able to put to use.
Theoretical numbers yes but the reality is not that far of.

But the lower the latency the better. That is what i was hoping for.

 

[myocardia]

Originally posted by: William Gaatjes
"Though you can buy MB's today like that but they are mainly used in office systems and some video systems, But cant make it in the videogame or image rendering world, I think Via still makes a few."

I doubt that cause when i look inside any older or current console...
For the image rendering world motherboards with memory upgrades can always still be made. And for the enthusiasts afcourse.

What do consoles have to do with desktop computers? Do you realize that one does only one thing, and the other does many thousands different of things? BTW, it would cost you at least twice as much money to buy a motherboard with that cheap RAM in it, if they came with the RAM soldered to the board. Want to know why? It's because each manufacturer would need to now sell 3 or 4x as many motherboards, because of all of the different RAM speeds and latencies that they would be asked to offer, for all of the different types of uses that computers have.

But the lower the latency the better. That is what i was hoping for.

In that case, you should be hoping that they never start soldering RAM to motherboards again, because solder doesn't help latency, and RAM sockets don't hurt latency. You'd almost certainly be forced to buy the most expensive processor, though, to get RAM with low latencies, which is all that counts with RAM latencies-- the IC's that are used.

 

[William Gaatjes]

What do consoles have to do with desktop computers ?

Well, if i have a look at for instance the PS3. The linux community is doing much effort of making this the ideal homecomputer. It does everything you want and has most of the connectivity a normal pc also has. I remember a first generation xbox which was nothing more then a trimmed down pc with on board vga and a special bios.
Not much difference.

As i have seen my pc does not do much then calculate. It is the addons like for instance printers which happens to connect to pc's that make it more alive. A usb burner buyable. I mean to say comparing the current consoles with pc's. not much difference with the exception of more memory and upgradability. And that was what my first post was about. When i upgrade , i usually upgrade all at once or at least mainboard, cpu and memory together. Afcourse the gfx card or HDD is a different story.

With exception of the power user with exotic pci /pci express cards most of the connectivity
is already on you board. Not much difference with a console. CPU, GPU, AUDIO, NETWORK, USB, WIFI, SATA, MEMORY all on 1 pcb. I have to add that the gpu is seperate afcourse on the pc architecture. And AMD 780 chips seem tempting. My personal dream is a more desktop oriented cpu like penryn but with IMC and (QPI or HT) connected to a AMD780 like chipset. And the memory soldered on the pcb if it gives me a win.


That is where i see a lot of parallels.

BTW, it would cost you at least twice as much money to buy a motherboard with that cheap RAM in it, if they came with the RAM soldered to the board. Want to know why? It's because each manufacturer would need to now sell 3 or 4x as many motherboards, because of all of the different RAM speeds and latencies that they would be asked to offer, for all of the different types of uses that computers have.

I don't think so and disagree. Why do we have different memories ? Because we have different generations of memories. I do not see you buying and using sdr memory or edo ram because it is sufficiënt in speed and bandwith for the pc you want to build. No you buy the memory that is currently available with the best price performance ratio. and currently that is ddr2. And it will be ddr3. And then another generation.

Most MB do not even provide support for more then1 memory generation with a few exceptions afcourse.
On cpu memory controllers make it more easy. Look at socket939 AM2, AM2+ and later AM3.
And now Intel is coming to with onboard memory controllers.


And with memory of different speeds is nothing special. It is just binning. Slower memory has lower timings and cannot be sold at a premium pricetag.

The fastest ram memory module out there does also not share the the same price with the cheapest ram memory modules out there.
So afcourse the price of memory is included in the board price afcourse. The faster the more you pay. Just as with seperate modules and MB's.
No difference there...

And adjusting timings is easy. We do it now too in the bios do we not ?

In that case, you should be hoping that they never start soldering RAM to motherboards again, because solder doesn't help latency, and RAM sockets don't hurt latency. You'd almost certainly be forced to buy the most expensive processor, though, to get RAM with low latencies, which is all that counts with RAM latencies-- the IC's that are used.

Riddle me this, why do we want optic connections in the future ?
Do you really think that even when cpu and chip designers already have to worry about how they route their circuitry, that the connection between chips on the pcb does not matter? I am not going to put to much text but i will just say there something like inductance,resistance and capacitance. And at higher frequencies they start to become more and more important. Then even a socket starts to make a difference.

 

[Modelworks]

Someone is already working on memory that is shaped a bit like a cpu.
It has the same type socket as those used with socket 478. The heatsink and memory is integrated so that it goes up from the socket.
Takes about 4 square inches on the motherboard for 16GB of memory

 

[myocardia]

Originally posted by: William Gaatjes
And the memory soldered on the pcb if it gives me a win.

No, it doesn't, it gives you a loss, for the reasons already explained to you in my first post in this thread. Typing more words than I type doesn't make you right. Feel free to buy all of the PS3's you want to buy. If you buy enough of them, and combine their capabilities, they may be able to do all of the things my slower computer does.

No you buy the memory that is currently available with the best price performance ratio. and currently that is ddr2. And it will be ddr3. And then another generation.

When did this become a discussion of EDO, and the types of RAM that were around before you were born? And much more importantly, what does that have to do with the size, type, and speed of RAM that I want to spend my money on?

Most MB do not even provide support for more then1 memory generation with a few exceptions afcourse. On cpu memory controllers make it more easy. Look at socket939 AM2, AM2+ and later AM3.

You obviously need to do more research into IMC's, because this statement proves you know zero about them. BTW, what does an IMC have to do with RAM that's soldered to a motherboard?

The fastest ram memory module out there does also not share the the same price with the cheapest ram memory modules out there.

And which are you wanting the memory manufacturers to do, throw away the half oftheir IC's that can't run @ XXX Mhz, or just lose the profit they make from people willing to pay more for their RAM, driving UP the price of the cheap RAM you like buying? You can't have it both ways.

So afcourse the price of memory is included in the board price afcourse. The faster the more you pay. Just as with seperate modules and MB's. No difference there...

Read this carefully, and maybe you'll understand it this time:

Originally posted by: myocardia
You'd almost certainly be forced to buy the most expensive processor, though, to get RAM with low latencies, which is all that counts with RAM latencies-- the IC's that are used.

And adjusting timings is easy. We do it now too in the bios do we not ?

Wrong. It's called SPD, look it up.

Riddle me this, why do we want optic connections in the future ?
Do you really think that even when cpu and chip designers already have to worry about how they route their circuitry, that the connection between chips on the pcb does not matter? I am not going to put to much text but i will just say there something like inductance,resistance and capacitance. And at higher frequencies they start to become more and more important. Then even a socket starts to make a difference.

You obviously know nothing about motherboards. Go do some googling of things like "traces", then get back to me when you understand that whether or not there's a gold-plated socket there, the majority of the capacitance, resistance, and induction takes place in the traces, not in the socket. BTW, if you would like me to explain to you why someday CPU's will use optical interconnects, I'd be happy to. It has zero to do with processor speeds or latencies, though.

 

[Mr Fox]

Originally posted by: myocardia

Originally posted by: William Gaatjes
And the memory soldered on the pcb if it gives me a win.

No, it doesn't, it gives you a loss, for the reasons already explained to you in my first post in this thread. Typing more words than I type doesn't make you right. Feel free to buy all of the PS3's you want to buy. If you buy enough of them, and combine their capabilities, they may be able to do all of the things my slower computer does.

No you buy the memory that is currently available with the best price performance ratio. and currently that is ddr2. And it will be ddr3. And then another generation.

When did this become a discussion of EDO, and the types of RAM that were around before you were born? And much more importantly, what does that have to do with the size, type, and speed of RAM that I want to spend my money on?

Most MB do not even provide support for more then1 memory generation with a few exceptions afcourse. On cpu memory controllers make it more easy. Look at socket939 AM2, AM2+ and later AM3.

You obviously need to do more research into IMC's, because this statement proves you know zero about them. BTW, what does an IMC have to do with RAM that's soldered to a motherboard?

The fastest ram memory module out there does also not share the the same price with the cheapest ram memory modules out there.

And which are you wanting the memory manufacturers to do, throw away the half oftheir IC's that can't run @ XXX Mhz, or just lose the profit they make from people willing to pay more for their RAM, driving UP the price of the cheap RAM you like buying? You can't have it both ways.

So afcourse the price of memory is included in the board price afcourse. The faster the more you pay. Just as with seperate modules and MB's. No difference there...

Read this carefully, and maybe you'll understand it this time:

Originally posted by: myocardia
You'd almost certainly be forced to buy the most expensive processor, though, to get RAM with low latencies, which is all that counts with RAM latencies-- the IC's that are used.

And adjusting timings is easy. We do it now too in the bios do we not ?

Wrong. It's called SPD, look it up.

Riddle me this, why do we want optic connections in the future ?
Do you really think that even when cpu and chip designers already have to worry about how they route their circuitry, that the connection between chips on the pcb does not matter? I am not going to put to much text but i will just say there something like inductance,resistance and capacitance. And at higher frequencies they start to become more and more important. Then even a socket starts to make a difference.

You obviously know nothing about motherboards. Go do some googling of things like "traces", then get back to me when you understand that whether or not there's a gold-plated socket there, the majority of the capacitance, resistance, and induction takes place in the traces, not in the socket. BTW, if you would like me to explain to you why someday CPU's will use optical interconnects, I'd be happy to. It has zero to do with processor speeds or latencies, though.

Myo.... I respect your patience here .. I would not have touched it with a 10' pole....

 

[William Gaatjes]

No, it doesn't, it gives you a loss, for the reasons already explained to you in my first post in this thread. Typing more words than I type doesn't make you right. Feel free to buy all of the PS3's you want to buy. If you buy enough of them, and combine their capabilities, they may be able to do all of the things my slower computer does.

I am enjoying this .

Your pc is for you like a big car i guess

I did not wanted to make a comparision of absolute calculation power.
And i am not going too. My simple example was nothing more then to show that most people can do what they want to do with the calculation power of consoles. If a console packs more horsepower or less horse power is irrelvant. It pack enough punch, just like any entry pc . But i wll make it more easy for you to grasp : The mac mini is not that powerfull either but it can do everything most users want to do. That is my point. I am not talking about powerusers(whatever they may use it for games/editing, etcetera).

When did this become a discussion of EDO, and the types of RAM that were around before you were born? And much more importantly, what does that have to do with the size, type, and speed of RAM that I want to spend my money on?

It didn't. You are not seeing my point. Please read again.

You obviously need to do more research into IMC's, because this statement proves you know zero about them. BTW, what does an IMC have to do with RAM that's soldered to a motherboard?

I am aware of Intregrated Memory Controllers( for the peope following this discussion and wondering what it means).

I am also aware that manufacturers take caution that customers can not mix memory generations when that is not allowed. Do a google on that. Compare a lot of mainboards and you will see. I just noticed some parallels there.

And which are you wanting the memory manufacturers to do, throw away the half oftheir IC's that can't run @ XXX Mhz, or just lose the profit they make from people willing to pay more for their RAM, driving UP the price of the cheap RAM you like buying? You can't have it both ways.

I have never mentioned i was planning to buy the cheap ram. After research i always buy the more expensive type since this is usuall faster and the faster the memory, the less the cpu has to wait. That is what my original question was all about. Less wait states.

However, the average user does not care about speed. The lowest speed graded chips usually end up with users who just want a pc but do not need the absolute fastest speed.
Not a difference there...
And motherboards with memory sockets will continue to exist.

I am willing to pay for a fast quality board with fast quality memory as i have done before and will do again. I don't upgrade that often and when i do i want qualitity.

Wrong. It's called SPD, look it up.

Hahaha.

Guess what reads the contents of the SPD and loads that content into the configuration registers of chipset or cpu?

Do you think the memory SPD just uploads it's contents into the chipset or cpu itself ?
Please do a google.

You obviously know nothing about motherboards. Go do some googling of things like "traces", then get back to me when you understand that whether or not there's a gold-plated socket there, the majority of the capacitance, resistance, and induction takes place in the traces, not in the socket. BTW, if you would like me to explain to you why someday CPU's will use optical interconnects, I'd be happy to. It has zero to do with processor speeds or latencies, though.

i don't need to google it is part of my hobby and a little of my work .

Now you say it is in the traces. Before you say it is the chips. You are getting there .
You see tracelenght and layout is important. And guess what was my big question ?

If i solder the memory chips closer to the cpu i have less tracelenghts.

Current situation :
Cpu - coppertraces - MB connector - RAM pcb connector - coppertraces - ramchips.

My suggestion :
Cpu - coppertraces - ramchips.

I guess the best way to compare this is if somecould explain the timings of gfx cards.
And the ram timings of for example the XBOX360 since it uses dd3. The PS3 uses a advanced rambus memory which is maybe difficult to compare. Have to look on that.

My question again was How much difference in latencies it would make.


And about the optical questions ? Do a search on EEtimes.com and the IBM website.
Also realworldtech.com has some nice info about it.

A short conclusion :
When you use light for intercommunication, you don't have to worry about interconnect latency for a while.

There is less heat production. No crosstalk between signals. Don't have to worry as much about signal integrity. No electromagnetical interference. PCb design becomes simpler.

Afcourse for now, there are a lot of issues still to be solved.

 

[William Gaatjes]

To Modelworks.

That is interesting. Is there also a speed benefit ? It seems like that would be easier to cool then current memory chips. 1 big heatpipe and/or heatsink over the motherboard .

 

[William Gaatjes]

I found some research info on optical connections.


http://arstechnica.com/news.ars/post/20060918-7773.html

http://arstechnica.com/news.ar...dream-in-progress.html


http://www.realworldtech.com/p...icleID=RWT082807020032


And a startup company claiming it's technology will be used in the near future.

http://www.eetimes.com/showArt...VN?articleID=206800385

 

[Lorne]

Yeah, Wait, What?.

The only reason to surface mount is to save space and money This whole latency and crosstalk was and issue that was worked over 15+ yrs ago and has always be part of the designing scheme since, Your adding nothing new.
I wouldnt worry about crosstalk, Ive seen 140Ghz components work fine with no cross talk, But if your still worried you could always use silver for tracings but again its not practical due to cost.

Rambus memory, Your kiding right, Rambus memory latency is 3x what DDR is and with a cost of 4x of DDR its not practical and since DDR surpassed its bandwith back in 02 its been generally phased out.

Optical comunication on motherboards?, Thats nonsence, By the time you put the enough up/down converters on to actually run a simple C2D it would be the size of a desk and that not incuding the other needed stuff like HD and graphix.
Optical comunication is best for long distance transmission.

 

[William Gaatjes]

!40Ghz components yes. But at what power envelope ? What kind of material.

Most of the technology we see now and see introduced as new is around for more then 40years. For Example Seymour Cray and many others did all the developing already.
I agree with you.

You see optical transmission is already possible but the benefits do not outway the costs yet for wide acceptance. If you are willing to pay millions and millions of dollars then yes you can have anything you want.

The only reason to surface mount is to save space and money This whole latency and crosstalk was and issue that was worked over 15+ yrs ago and has always be part of the designing scheme since, Your adding nothing new.

Afcourse they do, but there is always a price to pay with those nasty laws of nature.

At this moment optic communication between chips is still in development fase.
When it breaks through and for instance IBM or Intel are able to mass produce this in a way that it is for the same price as copper traces, nothing stands in it's way to become a accepted technology.

The guys from rambus have been improving their technology too. Afcourse it is more expensive. It is not overproduced or at least used by the masses.

Optical comunication on motherboards?, Thats nonsence, By the time you put the enough up/down converters on to actually run a simple C2D it would be the size of a desk and that not incuding the other needed stuff like HD and graphix. Optical comunication is best for long distance transmission.

Do not take this personnally, but i have more faith in the technicians of , IBM, Intel ,HP ,Texas Instruments , Agilent to name a few. See the links for what they already capable of.


There already concepts of motherboards with waveguids (optic fibres) patented.
It is just a matter of time until it is a mass market technology.

A little more info, it is outdated, Intel is a lot further now.

http://www.deviceforge.com/articles/AT3588366215.html


But i still have not had any real answer on my original question.

 

[Modelworks]

Originally posted by: William Gaatjes
To Modelworks.

That is interesting. Is there also a speed benefit ? It seems like that would be easier to cool then current memory chips. 1 big heatpipe and/or heatsink over the motherboard .

I assume there is a speed benefit , it allows the memory to be placed closer to the cpu because of the smaller size . Closer to the cpu should allow for the interface to run faster since the trace lengths would be much shorter.

 

[myocardia]

Originally posted by: William Gaatjes
I am enjoying this .

I'm enjoying it as much as you. BTW, if you read something that I've typed that makes you think I was being rude or condescending, then reread it, because I'm doing my best not to be either, so it wasn't meant to read that way at all. You're very knowledgeable, compared to most new people to anandtech.

I did not wanted to make a comparision of absolute calculation power.

I understand that, and if we're talking about web browsing & e-mail, you're right. I use a 1.2 Ghz Celeron-M for web browsing, myself. It's fast enough for what 90% of the world uses their computer. My point was "what about the other 10% of the world?" That other 10% is what allows these companies to make a profit, while still selling you a 1GB stick of DDR2 for less than $20.

And I reread the portion where you mentioned EDO. What you fail to see is that while the majority of the world is buying DDR2 today, there are people buying not only DDR1, but also DDR3. And more importantly, for the past 30 years or so, computers have needed to have their total system RAM doubled every ~2 years. That's why we're at the point now where nearly all of us have 4GB of system RAM. Want to know how much system RAM my first computer had? It had 16 kilobytes.

I am aware of Intregrated Memory Controllers( for the peope following this discussion and wondering what it means).

I am also aware that manufacturers take caution that customers can not mix memory generations when that is not allowed. Do a google on that. Compare a lot of mainboards and you will see. I just noticed some parallels there.

Yeah, they do that with memory sockets. Also, in case you weren't aware of it, no company has ever built a CPU with an IMC that can control more than one type of RAM. Sure, one of them keeps making promises that they will, but they've made alot of promises over the last few years, and not one of them has come true so far.

I have never mentioned i was planning to buy the cheap ram. After research i always buy the more expensive type since this is usuall faster and the faster the memory, the less the cpu has to wait. That is what my original question was all about. Less wait states.

Fair enough. Do you also want to be forced into buying one of these three processors, so you can get either faster, or lower latency RAM? Because I don't.

Guess what reads the contents of the SPD and loads that content into the configuration registers of chipset or cpu?

Haahaa, I guess I didn't explain that well enough-- I was getting tired of typing. My point was, it's 100% impossible to make PC2-3200 (400 Mhz DDR) run @ PC2-9600 (1200 Mhz DDR) speeds, no matter what changes you make in the BIOS. You can't just "decide" you're going to run it @ 1200 Mhz DDR, like the PC9600 is guaranteed to do.

Now you say it is in the traces. Before you say it is the chips. You are getting there .

No, latency is in the IC chips, plain and simple. I said that resistance, induction, and capacitance takes place in the traces.

You see tracelenght and layout is important.

Yes, it is. Why did you think that all motherboards now have the RAM sockets right next to the CPU socket? Were you aware that they used to (sometimes) put them at opposite ends of the motherboard? As memory speeds rise, it's more and more important that the RAM is as close as possible to the CPU, because of something called skew, not because of latencies.

If i solder the memory chips closer to the cpu i have less tracelenghts.

No, unless you solder them on top of the CPU, the tracelengths are the same.

My question again was How much difference in latencies it would make.

It would make no difference at all to latencies. That's determined completely by the IC's that are used, like I keep telling you.

And about the optical questions ? Do a search on EEtimes.com and the IBM website.

I have no questions about optical interconnects. I offered to explain them, remember?

When you use light for intercommunication, you don't have to worry about interconnect latency for a while. There is less heat production. No crosstalk between signals. Don't have to worry as much about signal integrity. No electromagnetical interference. PCb design becomes simpler.

I see you've done your homework about optical interconnects. You drew a wrong conclusion, though. Once you're using light for intercommunication, you've reached the epitome of speed. Nothing does or can go faster than the speed of light. All you'd be able to do after developing optical would be to improve the speed of the eqipment on either end of the light (the ending and receiving equipment, IOW).

It seems like that would be easier to cool then current memory chips. 1 big heatpipe and/or heatsink over the motherboard .

I agree. I would love to see the RAM mounted into a small, square socket, right next to the CPU, especially if they constructed it like a CPU socket, where a heatsink/heatpipe could be placed on it.

edit: Oh yeah, welcome to anandtech, William. You make a nice addition to the forums.

 

[myocardia]

Originally posted by: Modelworks
I assume there is a speed benefit , it allows the memory to be placed closer to the cpu because of the smaller size . Closer to the cpu should allow for the interface to run faster since the trace lengths would be much shorter.

You're right. If there's half the distance, there'd be roughly half the skew, presumably.

 

[William Gaatjes]

I'm enjoying it as much as you. BTW, if you read something that I've typed that makes you think I was being rude or condescending, then reread it, because I'm doing my best not to be either, so it wasn't meant to read that way at all. You're very knowledgeable, compared to most new people to anandtech.

Thank you.
Nice to see you are more up to date then i expected too.
And i do not want to be rude either. When i notice some one can follow me i do not have the desire to type it out for them. Call it lazy, i admit it.

I usually do not join forums for the simple reason that i do not want to participate in a yes / no discussion.

I find it hard to talk about these kinds of subject while the other person starts how they get 5 fps more over the 150 they already have. Those 5 do not really matter anymore i n my point of view.
It would if that person went from 25 to 30 fps for example. This is a real life example of for example on my work. I can remember when core 2 duo came out and read about it inner workings at arstechnica. I explained to a collega how good it was en he sad no non no pure shit. 1 month later on he starts bragging how good it overclocks and how fast it is like he discovered the lost ark or something. The exact same thing happened before that with AMD 64 with IMC. I sad it is a great cpu. I got the reply it was too slow, it did not overclock well. yada yada yada .
Later on ,it really did not have a need too compared to the P4, and he was bragging again . P4 by the way, had some great features too. However all those chips have tradeoffs making them faster in situation A and slower in situation B.

Boring. Alas , the rolling stones have a chorus line on that. You can't always get what you want.

I understand that, and if we're talking about web browsing & e-mail, you're right. I use a 1.2 Ghz Celeron-M for web browsing, myself. It's fast enough for what 90% of the world uses their computer. My point was "what about the other 10% of the world?" That other 10% is what allows these companies to make a profit, while still selling you a 1GB stick of DDR2 for less than $20.

I think , those 10 percent are important for making especially gfx speed ups. Most of the time i believe the oem's are the ones that make sure the bills of AMD and especially Intel get payed more or less. But, i think those 10 percent are the ones that make AMD and Intel try new things too. Afcourse in mine opinion , the memory manufacturers make enough money. Or they would not have overproduced.

Yeah, they do that with memory sockets. Also, in case you weren't aware of it, no company has ever built a CPU with an IMC that can control more than one type of RAM. Sure, one of them keeps making promises that they will, but they've made alot of promises over the last few years, and not one of them has come true so far.

Indeed , that was why i think it is possible to put the memory on the board. If it really has a measurable performance enhancement.

Fair enough. Do you also want to be forced into buying one of these three processors, so you can get either faster, or lower latency RAM? Because I don't.

I assume that all cpu IMC are guaranteed to work at given frequencies by the manufacturer. If not , it will be reveiled and the competition has a chance for new customers cause their cheaper cpu's can deliver. That is why there is always a strong feeling against monopolies.

Haahaa, I guess I didn't explain that well enough-- I was getting tired of typing. My point was, it's 100% impossible to make PC2-3200 (400 Mhz DDR) run @ PC2-9600 (1200 Mhz DDR) speeds, no matter what changes you make in the BIOS. You can't just "decide" you're going to run it @ 1200 Mhz DDR, like the PC9600 is guaranteed to do.

True. I thought you where bragging about having some knowledge about the spd .
Truly funny.




I did not draw the wrong conclusion, your last sentence says it all.

All you'd be able to do after developing optical would be to improve the speed of the eqipment on either end of the light (the ending and receiving equipment, IOW).

I think when the optical interconnects work , the engineers are very happy, they don't have to worry about the delay between interconnects anymore. When you design i/o interfaces , you design them with the communication path in mind(In this case coppertraces). That is where all the compensation is for and the delays that are necessary. Part of those delays exist as the chips latency. Turn around times, skew. I meant to say that with current technologies the chip can output data faster then coppertraces can follow without exotic expensive materials or serious power draw. It is just delays too. And the less delays , the better.


Afcourse you are right too. Accessing memory pages cost time. No doubt about that.


I think of a possible future when there is less heat from the i/o circuits cause they are optic, more interleaving/ banking cicrcuitry is possibe. The dramcells do not really become that much faster it seems, just bigger in numbers. Every generation there are just working more parallel, keeping more pages and/or sense amplifier rowes open and more logic to manage all those open pages.

It would make no difference at all to latencies. That's determined completely by the IC's that are used, like I keep telling you.

See above.

I have no questions about optical interconnects. I offered to explain them, remember?

I misread that, your quite right.

 

[William Gaatjes]

And I reread the portion where you mentioned EDO. What you fail to see is that while the majority of the world is buying DDR2 today, there are people buying not only DDR1, but also DDR3. And more importantly, for the past 30 years or so, computers have needed to have their total system RAM doubled every ~2 years. That's why we're at the point now where nearly all of us have 4GB of system RAM. Want to know how much system RAM my first computer had? It had 16 kilobytes.

I cannot control myself


Atari 600XL.


ME TOO !



I understand...
I bought some ddr for my brother while ddr2 was lot's cheaper.

But the same story goes when the memory is soldered on the board. Even when IMC's can handle different memory generations.
How many AMD chipsets are there now for example?
Since The Athlon 64 , the chipset is more of a peripheral controller.

And the 780 is something really new again with the possibility of the igp to help the seperate gfx card. Or let the igp(with it's own local memory) do some real physics and not just visual physics. Leave the graphics to the pci-express card. Many possibilities.
MP3 codec running on you igp or gfx for example through some vector api. Dreams, dreams...


I am wondering off....

 

[myocardia]

Originally posted by: William Gaatjes
Indeed , that was why i think it is possible to put the memory on the board. If it really has a measurable performance enhancement.

Yes, but just because something has a measurable improvement, doesn't meake it worthwhile to do. I'll use your earlier example of 150 FPS vs 155 FPS as an example.

I assume that all cpu IMC are guaranteed to work at given frequencies by the manufacturer.

Sure they are, but if RAM comes soldered to the motherboard, you aren't going to have the chance to use PC8500 DDR2 with your E4500/E2180. To be able to buy PC8500, or even 9200 or 9600, you'll have to buy the $500 motherboard (prices will be much higher, because of the RAM cost, remember). Then, what happens in six months, when you discover that 4GB wasn't enough? Well, you then have to buy the $650 motherboard (twice as much RAM= even more money). And if motherboard manufacturers start soldering RAM to motherboards, it wouldn't be long at all before all motherboards came with a CPU soldered to them.

I did not draw the wrong conclusion, your last sentence says it all.

Fair enough.

I think when the optical interconnects work , the engineers are very happy, they don't have to worry about the delay between interconnects anymore. When you design i/o interfaces , you design them with the communication path in mind(In this case coppertraces). That is where all the compensation is for and the delays that are necessary. Part of those delays exist as the chips latency. Turn around times, skew. I meant to say that with current technologies the chip can output data faster then coppertraces can follow without exotic expensive materials or serious power draw. It is just delays too. And the less delays , the better.

This paragraph made me rethink my position on latency. After more thought, it only makes sense that there's probably as much latency involved with getting the data back and forth from the CPU to the RAM as there is in the IC's themselves, and probably more, especially since you have to factor in the distance to the northbridge, add the northbridge's latency, then the "travel time" to the RAM IC's. That of course gets doubled, since the data has to make a return trip. Do you happen to know how long that roundtrip takes? I'd be interested in finding out.

I think that what Modelworks mentioned in his first post sounds very doable, and could make some serious performance differenes, if done correctly.

I think of a possible future when there is less heat from the i/o circuits cause they are optic, more interleaving/ banking cicrcuitry is possibe. The dramcells do not really become that much faster it seems, just bigger in numbers. Every generation there are just working more parallel, keeping more pages and/or sense amplifier rowes open and more logic to manage all those open pages.

Well, I'm not sure what the future holds (distant future). It seems very exciting to talk about it now, but by the time it finally happens, we'll all be saying "It's about time!", I can assure you. For instance, right now, I have a 3.2 Ghz quad-core, with 4GB of system RAM. If I had owned it 12 or 15 years ago, I could have rented it out for something like $200 million/year, because it would have been by far the fasest supercomputer on Earth (after I made the I/O faster, obviously). Today, it's not very extreme for a desktop, and in 5 years, it will be as slow as a Pentium III, compared to what we'll all be using.

Thanks for the discussion, BTW. If you keep posting here, I'll do my best to teach you a few things, and will also be happy to let you teach me what you know. Also, you should definitely browse the CPU forum, where I hang out the majority of the time. There are quite a few people there who know considerably more than I know.

 

[myocardia]

Originally posted by: William Gaatjes
I cannot control myself


Atari 600XL.


ME TOO !

I got a Tandy (Radio Shack) TRS-80 in 1980. Man, was I happy. I couldn't believe that my dad had bought it for me. I was the only kid in my school who had a computer, and was for a few years. As a matter of fact, alot of people either didn't know what computers were, or didn't realize that they no longer required their own building, like the only computer most of them had heard of, the NASA computer. I still can't believe it's been 28 years now. That means that I've owned a computer for twice as long as my daughter has been alive. That's funny.

 

[William Gaatjes]

A TRS 80 , with a Z80.

I have to be honest only seen pictures about it.


I started learning to program at school with a z80. Later i build some simple hardware for my msx2 i got later on and assembly programmed the msx2 to operate it. Just basic stuff about ad and da converters via I2C bus. Switching leds on and off and controling motors.

Now we have it easy with all those nice microcontrollers with onboard ad/da converters/timers etcetera.

 

[William Gaatjes]

Hello i think i have an answer to your question.

Do you happen to know how long that roundtrip takes? I'd be interested in finding out.

Well i am not sure i found the real world timings but in a datasheet of ddr2 dram I found they mentioned a a trw of 2 clockcycles. That is the time you have to wait minimal.

I was not sure if that was the only thing and if that timing is the only timing that has to do with the electrical characteristics of the bus. On this site they mention the same timings and give some explanation.

http://www.digit-life.com/arti...r2-rmma/ddr2-rmma.html

Interesting to know is that the internal dram cells of ddr run at 200Mhz while ddr2 dramcell run at 100 Mhz. DDR2 memory just picks up more data in parallel internally.
According to the website the 200Mhz to 100Mhz decrease dramcell clock has been done to improve yields.


I guess we will have to digg a little deeper.

How ddr3 works internally i don't know yet.

 

[William Gaatjes]

Could have editted the post above.

I think i shot myself in the foot.

It seems that the dram cell clock rises too when higher clocked i/o version come out.

The ddr2-400 was 100 MHz. The higher the i/o clock, the higher the clock of the memory cell.

I looked at a few datasheets but it is not that clear. Some manufacturers only mention 200 Mhz for the entire range, but that does not seem to make sense. If wikipedia is correct. memory cells clock can go up to 350 Mhz for ddr2.

DDR3 is not much different, same path : power consumption reduction, faster i/o and higher memory capacity.

 

[myocardia]

Originally posted by: William Gaatjes
A TRS 80 , with a Z80.

Yeah, it had a Zilog Z80. It was an 8-bit processor, and was considerably slower than the 8086, along with being an older design. Of course, I would have had to wait to get a computer with an 8086, because they weren't being sold in 1980 (at least in PC's, I'm sure they were probably in mainframes).

It seems that the dram cell clock rises too when higher clocked i/o version come out.

That's correct. You were right earlier, though, about the actual RAM speeds. DDR transfers data twice per clock cycle, DDR2 transfers data 4 times per clock cycle, and DDR3 transfers it 8 times per clock cycle. That's the reason that PC3200 DDR, PC2-6400 DDR2, and PC3-12800 DDR3 all run at the same speed, which is 200 Mhz. They just perform differently, because of how much data is being transferred, not because one is actually running at a higher frequency than the other(s).

 

[William Gaatjes]

Hello, I found this comment on realworldtech. My idea is pretty normal for a gfx card.

And i would not mind to buy a special Mainboard as i have explained with memory soldered on.
Afcourse while upgradable mainboards continue to exist.

2 options. Real fast non upgradable MB's and upgradable MB's with respect to memory.


I quote from the realworldtech site :

anon3 (no@spam.org) on 5/10/08 wrote: ---------------------------
>>No it won't. GPUs are a dying breed, invest in them at your own peril.
>>
> >Why is GPU dying? You're not going to get the best performance when GPU get integrated
>in the same package, say with a CPU. The physics won't allow it.
Are you just being >cranky today, or is there something only you got a head up?
> (Damn -- after re reading, I have to post anon -- Not violating any of the many NDAs I'm under, but possibly stepping on some toes none the less.)

I usually agree with Groo -- and on some of his above points, I still do. (BTW Groo, he probably meant PCI-express, not PCI-X)

Anyways, discrete GPUs are not going to die off any time soon. The reason is simple -- they are insatiable when it comes to memory bandwidth -- and having the GPU on a seperate chip gives you more pins to devote to GPU memory. Further, the requirements for GPU memory and main system memory are converging, but are still somewhat different -- mainly CPU memory needs to be easily expandable, and GPU memory does not - this affords soldering GPU memory directly onto boards, keeping the trace length down etc -- all things that make it easier to clock that memory much faster than CPU memory. Itegrating this onto one chip is possible (or will be) when you just consider the transistor count, but the pin count will be unfeasible -- multiple ddr3 channels and a 256 or 512 wide GDDR5 in the same package along with all the power, ground and other IO pins? Not bloody likely in the next 10 years.


On the other hand, the link between a sperate CPU and GPU will be getting faster with CSI, or whatever the hell Intel is calling it these days.

What may happen is that dedicated makers of discrete GPUs with undocumented internal architectures and instruction sets (yes ATI/NVidia, I'm looking at *you*) could have their days numbered.
More and more people like me want to be able to use the GPU for general number crunching. While CUDA is a small step in more or less the right direction, it is not nearly enough.
Intel has (or more accurately "will have") a real advantage here -- with making Larrabee an (extremely parallel) x86 derivative, their rich tool chain and dev libraries (think along the lines of TBB and others) will come into play, and will be very hard to compete with or match. The attraction of using the same programming models *and* tools on the GPU that I use to thread code on multiple CPU cores is very, very attractive. That is what Larrabee promises to bring to the table, and that is what could seriously hurt or kill NVidia.
(Not to mention the performance advantage Intel will be able to open up and maintain if they manufacture it on their latest fabs.)

Also they talk about the number crunching capabilities of gpu's for more general use.


Here is the link. Pretty fascinating.


http://www.realworldtech.com/f...hreadid=90000&roomid=2