AMD CPU speculation... and expert conjecture

[Cazalan]

According to the rumor, TSMC and GloFo are the two most likely bidders; who else is in the Foundry business these days? Of course, GloFo probably can't even afford IBM's foundry's, so that leaves TSMC as the only possible buyer.

GF could afford it. They have some really deep pockets until the oil fields run dry.

 

[con635]

http://wccftech.com/amd-kaveri-dual-graphics-benchmarked-frame-pacing-fix/
Read this and was wondering does the 'fix' cover the older trinity apus and hd 6670? Some nice results there the 7600 and r7 240 ddr3 looks a good budget/sff build.

 

[Cazalan]

Dual Graphics finally with frame pacing fix. Looks good.

http://wccftech.com/amd-kaveri-dual-graphics-benchmarked-frame-pacing-fix/

 

[con635]

lololololololol^^^^^^ great minds mate, great minds!

 

[8350rocks]

@juanrga:

http://semiengineering.com/whats-10nm/

http://www.advancedsubstratenews.com/2013/06/which-will-hit-the-14nm-jackpot-first-fd-soi-or-finfet-gauntlet-down-race-on-2/

http://www.electronicsweekly.com/mannerisms/manufacturing/arm-fancies-fd-soi-2013-07/

What was that about FinFET on bulk again? Just wondering...because I cannot find anyone outside Intel using FinFETs, and even they concede they need a new trick pony past 10nm.

As, I said...FD-SOI is the superior technology, and density was not an issue, it was the actual fabs mastering the tech.

 

[juggernautxtr]

lololololololol^^^^^^ great minds mate, great minds!

wonder what board they using, that looks absolutely awesome, may have to get me that 7600 for my son, that chip is stupendous.

 

[blackkstar]

@juanrga:

http://semiengineering.com/whats-10nm/

http://www.advancedsubstratenews.com/2013/06/which-will-hit-the-14nm-jackpot-first-fd-soi-or-finfet-gauntlet-down-race-on-2/

http://www.electronicsweekly.com/mannerisms/manufacturing/arm-fancies-fd-soi-2013-07/

What was that about FinFET on bulk again? Just wondering...because I cannot find anyone outside Intel using FinFETs, and even they concede they need a new trick pony past 10nm.

As, I said...FD-SOI is the superior technology, and density was not an issue, it was the actual fabs mastering the tech.

http://www.advancedsubstratenews.com/2013/07/globalfoundries-on-cost-vs-performance-for-fd-soi-bulk-and-finfet/

Yeah, looking for information about Bulk or SOI "winning" and GloFo is awfully ambiguous. The older sources all talk about how FD-SOI will be better cheaper and better performing. And then it suddenly fell out of the picture?

The way the GloFo articles from 2010 to current day play out, after looking at available information, leads me to side with you in that it was cancelled. Not because it was inferior technology, but because of other reasons, such as not enough customers, couldn't make it work, etc.

From what I've googled, the falling apart of GloFo post 28nm plans kind of went like this.

22nm SOI (cancelled) -> 20nm FD-SOI (cancelled) -> 28nm Bulk -> 20nm Bulk? -> 14nm bulk finfet?

My point is that something happened. The thing to remember is that by selling the fab, IBM will still be a customer of whomever buys it from IBM and there is room for another company to use it.

The obvious answer is AMD, but there are probably other companies who would consider it as well. It does seem like SOI is not so good for GPUs as it can't be as dense, so it puts things in a difficult situation as most products sold now are mostly GPU with a little CPU thrown on.

I wonder if FD-SOI reduced costs would make up for worse yields due to die size (if they yields are even worse at all) or not.

GloFo offering 22nm SOI would shake things up massively for products which are more CPU oriented.

EDIT: but I do wonder where all this bulk talk comes from. http://www.advancedsubstratenews.com/ is overwhelmingly posting about SOI and new partners and all sorts of things. Watching SOI developments seems to me like a solid thing to watch to see what AMD plans on doing in the future regarding more high endproducts that would work best on SOI.

 

[juggernautxtr]

@juanrga:

http://semiengineering.com/whats-10nm/

http://www.advancedsubstratenews.com/2013/06/which-will-hit-the-14nm-jackpot-first-fd-soi-or-finfet-gauntlet-down-race-on-2/

http://www.electronicsweekly.com/mannerisms/manufacturing/arm-fancies-fd-soi-2013-07/

What was that about FinFET on bulk again? Just wondering...because I cannot find anyone outside Intel using FinFETs, and even they concede they need a new trick pony past 10nm.

As, I said...FD-SOI is the superior technology, and density was not an issue, it was the actual fabs mastering the tech.

http://www.advancedsubstratenews.com/2013/07/globalfoundries-on-cost-vs-performance-for-fd-soi-bulk-and-finfet/

Yeah, looking for information about Bulk or SOI "winning" and GloFo is awfully ambiguous. The older sources all talk about how FD-SOI will be better cheaper and better performing. And then it suddenly fell out of the picture?

The way the GloFo articles from 2010 to current day play out, after looking at available information, leads me to side with you in that it was cancelled. Not because it was inferior technology, but because of other reasons, such as not enough customers, couldn't make it work, etc.

From what I've googled, the falling apart of GloFo post 28nm plans kind of went like this.

22nm SOI (cancelled) -> 20nm FD-SOI (cancelled) -> 28nm Bulk -> 20nm Bulk? -> 14nm bulk finfet?

My point is that something happened. The thing to remember is that by selling the fab, IBM will still be a customer of whomever buys it from IBM and there is room for another company to use it.

The obvious answer is AMD, but there are probably other companies who would consider it as well. It does seem like SOI is not so good for GPUs as it can't be as dense, so it puts things in a difficult situation as most products sold now are mostly GPU with a little CPU thrown on.

I wonder if FD-SOI reduced costs would make up for worse yields due to die size (if they yields are even worse at all) or not.

GloFo offering 22nm SOI would shake things up massively for products which are more CPU oriented.

EDIT: but I do wonder where all this bulk talk comes from. http://www.advancedsubstratenews.com/ is overwhelmingly posting about SOI and new partners and all sorts of things. Watching SOI developments seems to me like a solid thing to watch to see what AMD plans on doing in the future regarding more high endproducts that would work best on SOI.

I think AMD is going to start owning Intel from kaveri on, obviously the better gpu, the processor side is constantly improving, Intels architecture i think is peaking out, AMD is just getting the fx going.

it's going to depend on global foundries and or Tsmc, which ever finds the best process's to use to produce AMD's chips.

 

[jimmysmitty]

@juanrga:

http://semiengineering.com/whats-10nm/

http://www.advancedsubstratenews.com/2013/06/which-will-hit-the-14nm-jackpot-first-fd-soi-or-finfet-gauntlet-down-race-on-2/

http://www.electronicsweekly.com/mannerisms/manufacturing/arm-fancies-fd-soi-2013-07/

What was that about FinFET on bulk again? Just wondering...because I cannot find anyone outside Intel using FinFETs, and even they concede they need a new trick pony past 10nm.

As, I said...FD-SOI is the superior technology, and density was not an issue, it was the actual fabs mastering the tech.

It is not FinFETS that they need to lok over at 10nm, it is more the material they are using overall. When they went to HK/MG at 45nm, it was a jump before they really needed it but it along with hafnium allowed for much better temps at 45nm and much higher clock speeds. the QX9650 was insane when it came to overclocking and power use.

but beyond 10nm, they are starting to push the limits of the materials being used and will probably, or already are, be looking into new materials that will hold up and perform well at 10nm. FinFETs and 3D transistor technology will probably be much more prominent in that it will be used to stack and add more cores/cache to the same die space.

I think Intel is probably well on their way to looking into what materials are going to be needed at 10nm and beyond, not sure on GloFo though as they are still a bit behind Intel in process.

 

[de5_Roy]

kaveri has an unbeatable advantage over intel's mainstream cpus - full 64GB (4x 16GB) ram support. current intel core i5/i7 cpus can only support up to 32GB (4x 8GB).
http://www.anandtech.com/show/7742/im-intelligent-memory-to-release-16gb-unregistered-ddr3-modules
it made me sad. i really really wanted to max out my memory capacity with 4x 32/64GB ddr3 1600 modules. i <3 system memory. T_T

 

[juanrga]

It's always going to be heterogeneous as the primary model is Xeon + Xeon Phi. If you're just working on Xeon Phi then that is homogeneous.

By making it stand alone and removing the PCIe dependency they make it easier to deploy on cards with just a high speed ethernet port. It becomes much more scalable than the 4/8 or so PCIe cards you can fit in on a motherboard.

According to Intel, the primary model is standalone. The coprocessor version is released for legacy customers that already have Xeon CPU. Homogeneous systems are based around LCUs, not around TCUs.

@juanrga:

http://semiengineering.com/whats-10nm/

http://www.advancedsubstratenews.com/2013/06/which-will-hit-the-14nm-jackpot-first-fd-soi-or-finfet-gauntlet-down-race-on-2/

http://www.electronicsweekly.com/mannerisms/manufacturing/arm-fancies-fd-soi-2013-07/

What was that about FinFET on bulk again? Just wondering...because I cannot find anyone outside Intel using FinFETs, and even they concede they need a new trick pony past 10nm.

As, I said...FD-SOI is the superior technology, and density was not an issue, it was the actual fabs mastering the tech.

If you pay attention to the link that I gave you before, you can find answers to your questions.

As mentioned in the link, FinFET offers superior performance to FDSOI and scales down to 7nm, whereas there are doubts about SOI scaling beyond 10nm. This is the relevant part:

At 7nm there are differing opinions of the viability of FDSOI. There appears to be a clear path to 10nm FDSOI but 7nm is more controversial. At IEDM in December 2013 there was an FDSOI paper and the authors appeared to be confident 7nm could be achieved (keep in mind that the device layer has to get thinner as the gate length scales down). Other FDSOI experts I have spoken to are less optimistic.

The link gives a table with 14nm technology choices. Intel, Samsung, TMSC, UMC, and GF chose FinFET bulk. Only IBM chose FinFET on SOI, and only STMicro chose FDSOI (GF will also make FDSOI on demand under a manufacturing agreement with ST Micro). This account for 95% of the production being FINFET bulk and only 5% for SOI technology (FDSOI represents the 2.1%).

Now add the very recent news of IBM selling its foundries and SOI is almost dead as I predicted months ago. Your link about ARM is outdated. ARM gave a recent talk claiming leadership on FINFETs. ARM already tapped out test chip on 10nm FinFET bulk

And they are accelerating plans to get FinFETS on the market this year

http://www.electronicsweekly.com/mannerisms/manufacturing/16nm-finfet-arm-socs-market-year-2014-02/

You can find GF plans here

http://semiaccurate.com/2013/02/08/global-foundries-lays-out-their-finfet-plans/

 

[juanrga]

Another Kaveri review. This time it is vs Iris Pro desktop

http://wccftech.com/amd-kaveri-apu-a10-7850k-intel-i5-4570r-iris-pro-graphics-showdown/

This link is also interesting

http://www.techradar.com/news/computing/pc/amd-on-mantle-we-want-our-gaming-api-to-become-the-industry-standard-1218560

 

[GOM3RPLY3R]

It's always going to be heterogeneous as the primary model is Xeon + Xeon Phi. If you're just working on Xeon Phi then that is homogeneous.

By making it stand alone and removing the PCIe dependency they make it easier to deploy on cards with just a high speed ethernet port. It becomes much more scalable than the 4/8 or so PCIe cards you can fit in on a motherboard.

According to Intel, the primary model is standalone. The coprocessor version is released for legacy customers that already have Xeon CPU. Homogeneous systems are based around LCUs, not around TCUs.

@juanrga:

http://semiengineering.com/whats-10nm/

http://www.advancedsubstratenews.com/2013/06/which-will-hit-the-14nm-jackpot-first-fd-soi-or-finfet-gauntlet-down-race-on-2/

http://www.electronicsweekly.com/mannerisms/manufacturing/arm-fancies-fd-soi-2013-07/

What was that about FinFET on bulk again? Just wondering...because I cannot find anyone outside Intel using FinFETs, and even they concede they need a new trick pony past 10nm.

As, I said...FD-SOI is the superior technology, and density was not an issue, it was the actual fabs mastering the tech.

If you pay attention to the link that I gave you before, you can find answers to your questions.

As mentioned in the link, FinFET offers superior performance to FDSOI and scales down to 7nm, whereas there are doubts about SOI scaling beyond 10nm. This is the relevant part:

At 7nm there are differing opinions of the viability of FDSOI. There appears to be a clear path to 10nm FDSOI but 7nm is more controversial. At IEDM in December 2013 there was an FDSOI paper and the authors appeared to be confident 7nm could be achieved (keep in mind that the device layer has to get thinner as the gate length scales down). Other FDSOI experts I have spoken to are less optimistic.

The link gives a table with 14nm technology choices. Intel, Samsung, TMSC, UMC, and GF chose FinFET bulk. Only IBM chose FinFET on SOI, and only STMicro chose FDSOI (GF will also make FDSOI on demand under a manufacturing agreement with ST Micro). This account for 95% of the production being FINFET bulk and only 5% for SOI technology (FDSOI represents the 2.1%).

Now add the very recent news of IBM selling its foundries and SOI is almost dead as I predicted months ago. Your link about ARM is outdated. ARM gave a recent talk claiming leadership on FINFETs. ARM already tapped out test chip on 10nm FinFET bulk

And they are accelerating plans to get FinFETS on the market this year

http://www.electronicsweekly.com/mannerisms/manufacturing/16nm-finfet-arm-socs-market-year-2014-02/

You can find GF plans here

http://semiaccurate.com/2013/02/08/global-foundries-lays-out-their-finfet-plans/

Hi guys! Back again! The topic of discussion seems to be pretty interesting, so I thought about learning from you guys about this new stuff everyone is talking about. I also found a pretty interesting article on FinFET:

The innovation is based on a new transistor design called the FinFET because the transistors have the shape of tiny fish fins. TSMC’s new FinFET version is known as the "Omega FinFET", because the gate “wraps” around the silicon material that makes up the source and drain for each gate, creating a structure similar to the Greek character, Omega (W). Because of its novel design, the Omega FinFET operating at 0.7 volts nevertheless achieves a gate delay of 0.39 ps for the N-transistor and 0.88 ps for the P-transistor, making them among the fastest devices ever reported at this gate length.

http://www.tsmc.com/tsmcdotcom/PRListingNewsAction.do?action=detail&&newsid=1357&&newsdate=2002/12/11

I think other than the performance aspect of FinFET vs. FD-SOI which seems to have not completely been established yet, the 0.7 volts this ran at was interesting. I'll try to dig up more on it. ^_^

 

[ColinAP]

It would be interesting to see some cost / benefit analysis of bulk vs SOI, and even for a die shrink. Considerations would be:

1) R&D costs for moving process
2) Margin per chip sold
3) Performance gap vs opposition and own company's current chips
4) Value of additional sales due to a higher performing chip
5) Is your target market value focussed or performance focussed, and the size of that market

For most of the market, it seems that the benefits of being on SOI are outweighed by the cost of it, the exception being IBM.

The 250kg gorilla in the room - that's Intel, folks - long since decided that bulk makes them the most money, and they make a shedload of money. But even they have put off opening their 14nm fab...

 

[truegenius]

kaveri has an unbeatable advantage over intel's mainstream cpus - full 64GB (4x 16GB) ram support. current intel core i5/i7 cpus can only support up to 32GB (4x 8GB).
http://www.anandtech.com/show/7742/im-intelligent-memory-to-release-16gb-unregistered-ddr3-modules
it made me sad. i really really wanted to max out my memory capacity with 4x 32/64GB ddr3 1600 modules. i <3 system memory. T_T

How you gonna use this much amount of RAM ? Fraps !
and they said $320 for 1x16GB
and afaik all am3+ boards support only upto 32GB total (mine support only upto 8GB total ) so it doesn't seems legit for am3+ atleast now

 

[Cazalan]

How you gonna use this much amount of RAM ? Fraps !
and they said $320 for 1x16GB
and afaik all am3+ boards support only upto 32GB total (mine support only upto 8GB total ) so it doesn't seems legit for am3+ atleast now

Hah yeah. I only put 8GB in my more recent build and that only gets taxed because both Chrome and Firefox quickly suck up 1-2GB of memory each. They keep saying they're going to get more memory efficient but they're like hoovers sucking up as much RAM as they can get.

 

[de5_Roy]

How you gonna use this much amount of RAM ? Fraps !
and they said $320 for 1x16GB
and afaik all am3+ boards support only upto 32GB total (mine support only upto 8GB total ) so it doesn't seems legit for am3+ atleast now

hehe, whatever i can find.

i think kaveri, with 4+8 core configuration can make good use for 64GB on the cheap. cheaper than am3+ motherboards, since you won't need a discreet gfx processor. some fm2 motherboards also support up to 64GB afaik. you can fit max. 32GB in a small pc with mobo like this:
http://www.gigabyte.us/products/product-page.aspx?pid=4745#sp

as for am3+, boards like this are supposed to support 64GB. i don't think any one has tested for 64GB:
http://www.asrock.com/mb/AMD/970%20Pro3/?cat=Specifications
memory support is highly vendor dependent, so ymmv.

 

[Cazalan]

According to Intel, the primary model is standalone. The coprocessor version is released for legacy customers that already have Xeon CPU. Homogeneous systems are based around LCUs, not around TCUs.

If the Phi CPU is used stand alone it is homogeneous as all the cores are the same. There is no separation of LCU/TCU as in the AMD/NVidia model because there is only one type of core. Rajeeb Hazra commented as such.

"Knight's Landing, Hazra said, "takes us back to the veritable homogeneous architecture, except that it does that with the benefits of many-core for very highly parallel applications.""

http://www.theregister.co.uk/2013/11/19/intel_says_bootable_knights_landing_cpu_will_be_a_game_changer/

It is very much like BlueGene/Q on crack, as opposed to the CPGPU model which currently dominates the Green500 list. Intel is betting that the on-package 3D stacked memory will push the performance/watt to the top.

 

[8350rocks]

@juanrga:

Your article about FinFET plans from GF was from a year ago. If I posted such outdated material you would inquire as to why I cannot find a more recent article talking about my point...(in fact, all the articles I cited were more recent than yours).

Therefore, I ask you to supply a roadmap for GF plans now.

 

[jimmysmitty]

kaveri has an unbeatable advantage over intel's mainstream cpus - full 64GB (4x 16GB) ram support. current intel core i5/i7 cpus can only support up to 32GB (4x 8GB).
http://www.anandtech.com/show/7742/im-intelligent-memory-to-release-16gb-unregistered-ddr3-modules
it made me sad. i really really wanted to max out my memory capacity with 4x 32/64GB ddr3 1600 modules. i <3 system memory. T_T

Problem is getting 16GB modules at DDR3. Right now there are none. The only thing they have 16GB modules in is ECC DDR3 for servers and from what I have been reading consumer grade DDR3 16GB DIMMs are not going to hit, more than likely it will only come with DDR4.

In all honesty 16GB is a bit overkill, 32GB is overkill and 64GB is insane. For most people who will use Kaveri, it is just way more than they will ever need.

It's always going to be heterogeneous as the primary model is Xeon + Xeon Phi. If you're just working on Xeon Phi then that is homogeneous.

By making it stand alone and removing the PCIe dependency they make it easier to deploy on cards with just a high speed ethernet port. It becomes much more scalable than the 4/8 or so PCIe cards you can fit in on a motherboard.

According to Intel, the primary model is standalone. The coprocessor version is released for legacy customers that already have Xeon CPU. Homogeneous systems are based around LCUs, not around TCUs.

@juanrga:

http://semiengineering.com/whats-10nm/

http://www.advancedsubstratenews.com/2013/06/which-will-hit-the-14nm-jackpot-first-fd-soi-or-finfet-gauntlet-down-race-on-2/

http://www.electronicsweekly.com/mannerisms/manufacturing/arm-fancies-fd-soi-2013-07/

What was that about FinFET on bulk again? Just wondering...because I cannot find anyone outside Intel using FinFETs, and even they concede they need a new trick pony past 10nm.

As, I said...FD-SOI is the superior technology, and density was not an issue, it was the actual fabs mastering the tech.

If you pay attention to the link that I gave you before, you can find answers to your questions.

As mentioned in the link, FinFET offers superior performance to FDSOI and scales down to 7nm, whereas there are doubts about SOI scaling beyond 10nm. This is the relevant part:

At 7nm there are differing opinions of the viability of FDSOI. There appears to be a clear path to 10nm FDSOI but 7nm is more controversial. At IEDM in December 2013 there was an FDSOI paper and the authors appeared to be confident 7nm could be achieved (keep in mind that the device layer has to get thinner as the gate length scales down). Other FDSOI experts I have spoken to are less optimistic.

The link gives a table with 14nm technology choices. Intel, Samsung, TMSC, UMC, and GF chose FinFET bulk. Only IBM chose FinFET on SOI, and only STMicro chose FDSOI (GF will also make FDSOI on demand under a manufacturing agreement with ST Micro). This account for 95% of the production being FINFET bulk and only 5% for SOI technology (FDSOI represents the 2.1%).

Now add the very recent news of IBM selling its foundries and SOI is almost dead as I predicted months ago. Your link about ARM is outdated. ARM gave a recent talk claiming leadership on FINFETs. ARM already tapped out test chip on 10nm FinFET bulk

And they are accelerating plans to get FinFETS on the market this year

http://www.electronicsweekly.com/mannerisms/manufacturing/16nm-finfet-arm-socs-market-year-2014-02/

You can find GF plans here

http://semiaccurate.com/2013/02/08/global-foundries-lays-out-their-finfet-plans/

I am not impressed with ARM as their CPUs are very basic and do not require a lot of die space so moving to 10nm is probably very easy for them compared to moving a much more advanced and complicated x86 CPU like Intel or AMD has.

I still doubt they are much more ahead of Intel as they claim, remember Intel works a lot of stuff behind the doors and we wont know until they are near launch.

How you gonna use this much amount of RAM ? Fraps !
and they said $320 for 1x16GB
and afaik all am3+ boards support only upto 32GB total (mine support only upto 8GB total ) so it doesn't seems legit for am3+ atleast now

Hah yeah. I only put 8GB in my more recent build and that only gets taxed because both Chrome and Firefox quickly suck up 1-2GB of memory each. They keep saying they're going to get more memory efficient but they're like hoovers sucking up as much RAM as they can get.

I went 16GB and am thinking of moving to 32GB just because I like to always have more. I had 2GB of DDR when 512MB was the normal, 4GB DDR2 when 1GB was normal and 16GB of DDR3 when, well 8GB is now normal but when I started that much it was 4GB.

Since RAM is so cheap, it is easy to throw a ton in there so I don't think we will ever be in a position where we will ever have too little. I bet in the next 3 years 64GB will cost as much as 16GB does.

 

[de5_Roy]

Problem is getting 16GB modules at DDR3. Right now there are none. The only thing they have 16GB modules in is ECC DDR3 for servers and from what I have been reading consumer grade DDR3 16GB DIMMs are not going to hit, more than likely it will only come with DDR4.

In all honesty 16GB is a bit overkill, 32GB is overkill and 64GB is insane. For most people who will use Kaveri, it is just way more than they will ever need.

yes, kaveri's target demographic won't need more than 8-16GB let alone 64GB. but, if someone wants to use that much ram, it's a nice option to have.
amd bundles dataram's ramdisk software with some kits.
http://www.amd.com/US/PRODUCTS/DESKTOP/RADEON-MEMORY/Pages/ramdisk-overview.aspx
64GB can easily host a system drive or a new game on ram.

 

[gamerk316]

I again point out that a 32 app maintains its 4GB Address Space ceiling, even on Win64 systems. So, given how most programs are still 32-bit, and you typically only use one or two of them at a time, anything over 16GB is kinda overkill. Even 8GB is pushing it. Hence why improvements in performance via adding more RAM drops off at ~6GB or so.

 

[logainofhades]

16gb is not overkill. I have to run a pagefile or I get low memory warnings with 8gb. :lol:

 

[-Fran-]

16gb is not overkill. I have to run a pagefile or I get low memory warnings with 8gb. :lol:

Same here, and I am with 16GB.

I think we had this RAM usage discussion a while back (here, I believe, lol) and even Toms put a great article on how much RAM is actually a good amount and what uses you can give it.

More than 32GB is great when you want to run a page file or swap partition on it. Specially for speed ups on memory intensive apps (imagery and video). I run a RAID 0 config, so having a page file is not THAT bad, but I'd love to live without one or putting it on my RAM.

For low end systems, the speed up of having more RAM, given you can put a page file in there, would be gigantic. So, I'd say systems with slow as crap hard drives would have the biggest upgrade with more RAM.

Cheers!

EDIT: Added idea.

 

[palladin9479]

Ok there is some confusion about NT's virtual memory subsystem, notable the pagefile.

Whatever you do, do NOT put the page file on a RAM Disk, it's counter productive and possibly dangerous to system stability.

In NT you will ALWAYS have a pagefile and windows will ALWAYS complain if it's too small to fit all reserved memory in, even if your have 128GB of system memory. This is one of the major differences between NT and Unix/Linux, the way they treat memory. Whenever a program starts up in NT it makes memalloc requests for the OS to provision space for them. What NT does is assign a region in the pagefile for those memalloc requests, this is done as a safety measure to ensure there is someplace to put the data if it arrives. Now notice I said "if", that's because memalloc is not storing anything, it's just making a reservation. Physical memory isn't actually allocated until a program goes to store data inside memory. If there isn't anything to store then no physical memory is used but the reservations are still kept for the pagefile. This model ensures that a situation never occurs where a program goes to write something to memory and there is no space for it and thus it crash's. And since all process's use this memory subsystem, that includes drivers and kernel process's too. A process is more likely to run out of address space (2GB for NT32) then it is to run out of page space.

Now for why you don't store it on a ramdisk. The service that creates the ramdisk also runs through this same memory subsystem. If you have 128GB of memory and section off 32GB of it for "RAMDISK", then memory subsystem will try to allocate 32GB of pagespace for that RAMDISK even if you have plenty of memory available. It should now be apparent why putting a pagefile on a ramdisk would be a very bad idea. The pagefile is holding the reservations for that same ramdisk and the system can start freaking out when you go to write to it.

This is where the entire idea of "swapfile should be system memory x2" came out.