If anyone knows who said "640KB of memory ought to be more than enough for anybody." Can you please email the link and person to me at Bsb215@kiwibox.com o.k. Thanx.

Ryu (#57),
I don't understand your not understanding? We're talking about two completely separate sets of memory -

(i) the up-to-4GB of memory on the PC montherboard, addressed by the north-bridge

(ii) the memory on the PCI card,

Potentially, the memory on the PCI card could be addressed completely independently to the motherboard memory, with no reference to the north-bridge at all.

Obviously, this implies that the chip on the PCI card (eg an i960) would have to have two separate sets of data and address lines -(i) one set to the PC's memory via the PCI bus and northbrdige as per normal, and (ii) another set to the on-board memory, which is addressed without reference to anything else on the PC, least of all the northbridge.

My point was, if they design the card with independently addressed local memory, there is (by definition) no overlap with the PC's 4GB memory space. Alanzilla replied that in principle that's OK, but there are practical difficulties that might mean that you're going to need to reserve some on-board memory space in any case, eg for double-buffer writes (although I don't really understand the double-buffer write issue!).

Clear as mud?
Roger

You lost me there, Ryu. Could you try that again?

I still find the situation somewhat thick.

The PCI bus and its devices can either be independent or dependent.

A bus matered device can read and write to memory directly without CPU intervention.

A non bus mastered device cannot read and write to memory directly without CPU intervention.

If it is a hub architechture the request goes the south bridge and then to the north bridge.

In a traditional architecture the device talks straight to the north bridge.

In both cases the Northbridge has to pipe data back to the PCI bus. With AGP cards doing DMA and DME and certain PCI video cards like the Voodoo2 actually requring CPU intervention to do memory reads and writes, I have to wonder how any PCI device with memory can be over the addressable range of the Northbridge and CPU.

Roger (#54), I think, for the most part, you're right. It really isn't necessary, and up to now, there has been no downside to doing it that way.

If you were to design the board that way, though, you'd still need some kind of FIFO or buffer memory area to store things (PCI data transfers still have less latency than SCSI writes, I believe). Also, you'd most likely need to double-buffer writes (just like they do with SCSI cards), so you'd probably end up with the same size chunk of memory being devoted to the RAID card--the only difference would be that it is in main memory instead of being on the card.

Ryu (#45),
I meant that if the card is truly using local memory, it's only limited by the address range of the on-board chip (eg i960 or whatever), not by any of the architectural constraints of the processor and/or northbrigdge/MCH.

Roger

#39 again.

Alanzilla, thanks for the explanation. I understand your point (2), but I'm not clear on why the RAID card's driver needs to access the PC's memory in order to "manage multiple reads and writes instead of having to do one at a time"? (This is just me being a dummy, not me challenging your statement!)

Surely, if it was designed to do so, it could still cache the reads/writes in its local memory, while sending data to/from the shared memory along the PCI bus as and when required? Maybe up until now it's just been easier to use the shared memory, e.g. they only need one set of address/data lines instead of two (one set for local memory, one set for the PC's shared memory over the PCI bus).

Cheers,
Roger

BTW - That 3gb of Crucial PC100 Registered ECC set me back a whopping $512 ;-)

AG #52

Hehehe... I thought the same thing! That's why I recently added another 3gb of memory to my Exchange server here at work.

I figured that 4gb of memory would go well in the quad 700/1mb Xeon machine with 252gb of 10k SCSI RAID. I think Exchange should be happy for some time to come.

it would be pretty interesting, if over the next couple years, RAM configs tend to "plateau" aroung 4gig, just like it did around 640K/1MB back in the day.

Oh, and the "remember when RAM cost a buck a meg?" comment made me feel really old when I could remember it being $75 a meg, when I was in high school.

Oh, my tapioca's calling.

ER

#49. Id dont think you quite have a grasp of chip architecture. 64 bit will be necesarry jsut for the ram at first but of key importance is moveing away from the x86 ISA. Actually Itamium is quite fast in certian situations and offers great promise for future generations. x86 poor ability to scale in performance is due to its command set. While an EPIC or RISC system more ipc can be extracted from the code and thus wider issue becomes advantageous. X86 will eventually need to go away. I think Intel figured that it should go away during the shift to 64bit as that will require a recompile anyway so why not just have them recompile to a differnt ISA.

Amds hammer line allthough nice puzzles me. It seems it would be a niche product. Well the market may decide they like it from a performance/price standpoint. but from an engineerig standpoint its been designed for a specific niche. That niche being cheap servers to run legacy apps that require more memory than 4GB. All in all this dosent make alot of sense to me as an engineer as you will have to implement the instructions so a recomplie is necesary so why not recompile for a better ISA. All in all I guess I dont understand AMD's reasoning behind the architecture. Intel will be selling a high end server chip and a low end desktop chip/s. The only reason I can think of for hammers design is AMD wants to sell many different SKU's all based of of one core architecture. thats jsut my resoning though. Any other EE people out there have different opinions?

I still think > 4gb of RAM is worthless for 99.99% of people, and will remain so for some time. We don't need that much RAM for most anything that most people do. The extra space consumed in addresses will in fact generally [ilow[/i] us all down. Think efficiency!

The Hammer is really the [berfect[/b] 64 bit mainstream chip, that is to say: it runs in 32bit 99% of the time, and is perfectly compatable with all the current software. Sure its a hack: x86 is ugly. Thats why they have high-level languages and operating systems, so who cares if we still have x86 in 2040? I am rooting for AMD's course, screw Intel's goofy new ideas.

Can anyone imagine an Itanium in a server? Wouldn't that be slow?

Ryu (#45), not really, no. It's a function of how the card itself is architected.

The typical RAID card that I deal with has a PCI bridge with an i960, RAM, and a Qlogic SCSI chip behind it.

The architecture has its disadvantages, of course, including the using-up-space-in-main-memory issue, but it also has some advantages.

If you have more than one card in the machine at a time, they can operate in "ganged" mode, since they all can talk to each other via the PCI bus. Likewise, they only run one option ROM which handles all the cards. Also, if you have an extra SCSI card in the system, the RAID card can include the drives from the SCSI card in RAIDgroups with its own drives.

As for why the RAID card's own memory is available in main memory, well, there are two pieces to that. The reason the memory exists at all is to serve as a big buffer for disk transactions. The two reasons it is in main memory:
1. The driver: the RAID card's driver can manage multiple reads and writes instead of having to do one at a time (or double-buffering like many of the SCSI card drivers).
2. Tradition: up until just very recently, you didn't see machines with more than 2GB, so nobody worried about taking big chunks of memory up near the 4GB limit.

Some chipsets provide for the ability to move memory hidden by the PCI hole above the 4GB limit, but most don't. I'm currently working on a BIOS solution for this issue, actually, but there are lots of little side-effects which may make it not worth it.

Interestingly, 98/Me have an even lower limit: 1GB. Giving a machine 1.5GB or more can cause some neat "infinite-reboot" issues. MS posts a workaround that lobotomizes anything over 1GB, allowing multi-booters to keep going.

http://support.microsoft.com/support/kb/articles/Q304/9/43.ASP

ScottG

Not case sensitive. Just watch out for guys called 'F0RGE' or 'FORG3'.

[q]#39, I'm not precisely certain why PCI cards map their memory into main memory. Most of the information that I have does not answer that satisfactorily. I'll do some research.[/q]

Wouldn't that be a limitation of the processors and northbrigdge/MCH servicable address range?

Originally Posted by localyokel
Forge, is that really you, or are usernames case sensitive?

#39, I'm not precisely certain why PCI cards map their memory into main memory. Most of the information that I have does not answer that satisfactorily. I'll do some research.

oh and I forgot to mention Windows XP incorporates an even better system chache algorithm that uses prefectching from thee hard drive so more ram will be even better then.

Let me clarify some points for some of you.

Raid cards msut map in to the 32bit adress space as an ugly hack but its necessary due to the even smaller io adress space. You couldnt even fit 512MB in the iorage. So mapping becomes necessary.

Secondly, You can not use task manager to determine the amount of ram that is currently ocupied in your system. If you have acess to a PCI disgnostic board you can see that windows uses allmost all of the ram instantly. Particularly if you select the LARGE SYSTEM CACHE registry value. This value will substantially increase boot time if you have large amounts of ram but thats because it atempts to load all system dll files in to ram and then offers the rest up to the cache.

Secondly 32bit Windows Apps are not limitedto 2GB or ram. You are right in your understanding of the memory model but not in the implementation of its uses. If more than 2GB of ram is necessary programmer do one of two things. First being to request system heap space. This works well and is fast but is not multiprocessor freindly so if you are making a multithreaded app wich this will probobly be considering the memory requirements puts it in the server area where multicpu is common. than you spawn a new process or com object and comunicate with it in a variety of ways. Also note that jsut because an app has 2GB of system adressable data also does not mean it cant use more than that. Yes 64bit is the way to go but its certianly no hurry.

In my mind you are all missing the key advantage of this dirt cheap ram. When PCI-X finally gets here or even 64bit pci think solid state disk storage. Check out platypus line of solid state ram drives. I bought one of these and its simply amazing. Talk about performance increase. The only problem is my PCI bus is completely saturated most of the time. Hopefully that will be correct soon with my purchase of a Tiger MP wich offers two independent PCI buses.

Originally Posted by glutzie
Hey I only post the 98 mem crap to warn 98 users that more isn\'t always better. Just had someone come ask why they are running out of memory. Come to find out he just up his ram and prior he never ran out. Same shit all the time. I know this is on the low tech end here (or 98% of the home market).

Hell I have 512 MB but I also run a dual boot of 98/W2K (just for my incompatable progs).

Umm what more do you need in proof if when you add more mem and all of a sudden you start running out of resources?

This will all be moot when the great M$ bestows XP to all and then then revolution from 32 to 64 bit processing starts next year (maybe - hell nothing seems to happen on time).

Alanzilla said "This isn't too big a deal if you have a small RAID card (maybe 32MB of memory), but it is not uncommon for customer configurations to have five RAID cards with 256MB each. 5 * 256MB + a fudge factor for allocation routine inefficiency often means your 4GB of memory is really more like 2.5GB! "

Why does the card's memory have to be mapped to the PC's 4GB space, i.e. why can't the card could keep that memory privately in it's own address space, and access it as and when required in the course of reading/writing the disk?

Originally Posted by forge
Cripes....

Upcoming hardware on my lists:

Radeon MAXX. 2 * 128MB AGP devices.
Tekram SCSI card. no cost.
Big kahuna SCSI RAID card with 256MB cache.
2 GB system ram.

That\'s 2.5GB right there... I might outrun everybody to the 4GB barrier, but I already have PAE36 support on, so I\'d imagine we will all actually have to deal with the problem at the same point, years from now.

Originally Posted by LocalYokel
I doubt that any conclusive answer to the Win9x memory size issue, if it even exists or is the real problem, will ever be found. It\'s a complex issue with too many conflicting analyses and stories to ever produce a conclusion that everyone will accept.

May I suggest that we put this in the \"unsolvable\" file, next to \"Darth VIA\" and IBM GXP drives?

ummmm you SURE you can't use PAE on Pro or Server?

It's been a while but I could swear I've used the /PAE switch on a pro and/or server machine at some point or another..

Also for the most part the Kernal on all Versions is pretty much the same. They just have different regstry settings on boot that get loaded... SOOO I would think it wouldn't be to hard to hack the registry and make PAE workable on a "lesser" version of Windows..

I could be completely wrong as it's been a while since I've used PAE stuffs...

Don't matter to me anyway... I have a fully licensed LEGAL copy of 2K ads...but also i'm not even close to 4 gigs in any of my machines here..

Later

"And don't waste your money on over 64 MBs if your using anything less than W2K (ie: 98 & Me). Unless you don't mind rebooting often. Hell my friend is a big Chatter and he couldn't run his bots anymore, had to lose the memory."

I'd rather reboot my PC often than waste my time waiting for Photoshop to open or listen to my hard drive grind away because I ran out of physical memory while playing a game.

Originally Posted by mac_bug
Alanzilla, I love you man

LETS NOT GO THERE K WE DONT NEED TO TALK ABOUT THAT KINDA RAM

[q]LiamC (#24), PCI memory mapping is an interesting thing that most people don't usually run into, since server RAID cards are one of the biggest culprits in causing these situations.

A RAID card is usually just a SCSI card, an embedded processor (often some i960 variant), and some battery-backed memory. This memory is usually mapped to just under the 4GB limit, but nobody notices, because there usually isn't any physical memory up there. This works just fine, because the processor's Memory Type and Range Registers (the infamous MTRRs) just list physical memory and everything else is marked as "uncacheable" (which almost always means PCI space, nowadays) while the PCI northbridge looks for addresses way up there.

In the case that physical memory exists where the PCI device's memory needs to sit, the PCI device wins, and that area in physical memory just sits there unused because there's no way to access it. In other words, the MTRRs show a chunk up there that is marked "uncacheable" and the PCI northbridge is programmed to own that chunk. When a memory write or read occurs to that chunk, the northbridge grabs it, and physical memory is never touched.

This isn't too big a deal if you have a small RAID card (maybe 32MB of memory), but it is not uncommon for customer configurations to have five RAID cards with 256MB each. 5 * 256MB + a fudge factor for allocation routine inefficiency often means your 4GB of memory is really more like 2.5GB!

I've been meaning to work up some articles about stuff like this for a long time.... Damage has probably given up on them....[/q]

Alanzilla, I love you man. That was great.

It makes sense that the address range of the expansion bus would cause overlap in the adressable range of memory if you hit the maximum capacity. I had never considered the implications of such an action though.

Given how big the memory size of video cards are and the rate of growth for that memory, I suppose that they will soon rival your RAID cards. This memory issue may end up getting more recognition in the near future.

I for one, have 92MB of addressable memory on PCI cards just in my home box.

Actually, my characterization of the MTRRs isn't quite right. The PCI memory hole, as it's often called, is marked "uncacheable", main memory is marked "cacheable" and everything else is marked as "does not exist".

If you were to mark memory that doesn't exist as "uncacheable" you could cause a lot of problems, I'd imagine. Hmmm. Maybe I'll go try that and see what it does....

LiamC (#24), PCI memory mapping is an interesting thing that most people don't usually run into, since server RAID cards are one of the biggest culprits in causing these situations.

A RAID card is usually just a SCSI card, an embedded processor (often some i960 variant), and some battery-backed memory. This memory is usually mapped to just under the 4GB limit, but nobody notices, because there usually isn't any physical memory up there. This works just fine, because the processor's Memory Type and Range Registers (the infamous MTRRs) just list physical memory and everything else is marked as "uncacheable" (which almost always means PCI space, nowadays) while the PCI northbridge looks for addresses way up there.

In the case that physical memory exists where the PCI device's memory needs to sit, the PCI device wins, and that area in physical memory just sits there unused because there's no way to access it. In other words, the MTRRs show a chunk up there that is marked "uncacheable" and the PCI northbridge is programmed to own that chunk. When a memory write or read occurs to that chunk, the northbridge grabs it, and physical memory is never touched.

This isn't too big a deal if you have a small RAID card (maybe 32MB of memory), but it is not uncommon for customer configurations to have five RAID cards with 256MB each. 5 * 256MB + a fudge factor for allocation routine inefficiency often means your 4GB of memory is really more like 2.5GB!

I've been meaning to work up some articles about stuff like this for a long time.... Damage has probably given up on them....