K8 microcode editing to use 1xx opterons in 2+ socket motherboards. (This may be big)

[ScHpAnKy]

A good hypothesis is much more than a guess.

Discoveries are made by eliminating the unlikely first. There's very little to go on here, so everything is possible -- which makes the set of things to test impossibly large to complete by brute force.

Do you see a method for discovering the microarchitecure or microcode map that's more effective than brute force? What's your hypothesis for the structure of the microarchitecture, or the mapping of the microcode ops?

A hypothesis, by definition, is merely an educated guess
I was merely over-simplifying the procedure, but rudimentarily it is, indeed, Test and Guess

Do I see a method? This can be likened to asking if I see the answer, the obvious reply is that no, I do not, because I have not given it any valiant effort, nor the appropriate time and/or brain power. The microcode provides exciting potential, and I believe their is reasons for AMD to not have this released until this late in the game, but do I believe that it is going to help us run 939x2? Doubtful, simply because we need a motherboard to house the processors in the first place.

Sorry for making you read all that^

 

[Dan_D]

Well, AMD's latest headache inducing scheme kind of throws a wrench in that.
E4 (90nm 1MB) 939 Opterons are ECC Registered required, just like all Opterons. So yes, 939's can do ECC Registered, if they're E4 Opterons. And I'm pretty sure an Athlon64 will POST with ECC Registered, just won't use ECC. I don't have any A64's to check that with though.
Not that it does any good, but figured I should point that out. If AMD gets to cause me headaches, they get to cause you headaches too.

I actually have tried using Registered ECC memory with an A64 by mistake once. An Athlon 64 will not even post with it. The boards can use ECC memory, but not Registered ECC which is different.

As far as I know, all S939 Opterons are the same as their Athlon 64 and X2 counterparts, reguardless of stepping and revision.

 

[mikeblas]

Jesus fucking Christ, Mike, if you would like me to say that I've made a mistake in posting this, fine. My mistake.

Huh? Where the heck did that come from? I've harbored no such thought, and you must have wildly misinterpreted something.

A hypothesis, by definition, is merely an educated guess
I was merely over-simplifying the procedure, but rudimentarily it is, indeed, Test and Guess

You're skipping a very important step: analyse. And an educated guess includes insight, direction, and reduction -- which indeed make it substantially more than just a guess.

Do I see a method? This can be likened to asking if I see the answer,

Hardly; there are many problems where we know the method for solving them but don't know the answer.

the obvious reply is that no, I do not, because I have not given it any valiant effort, nor the appropriate time and/or brain power. The microcode provides exciting potential, and I believe their is reasons for AMD to not have this released until this late in the game, but do I believe that it is going to help us run 939x2? Doubtful, simply because we need a motherboard to house the processors in the first place.

I think the reason that it's obvious you don't is because the problem, without any insight (eg, leaked design documents from AMD) is NP complete. We know nothing about the microarchitecture's structre, and nothing about the opcodes it uses. We've got no tools to write (encode) or read (decode), test, or monitor what's happening. We don't even know if these features are controlled by microode, or if the on-chip hardware is there. And as you point out, we don't know what the existing motherboards and BIOSes would do if the support turned out to be enabled against chipset that didn't expect it.

Certainly, you must not really think that having a method to approach such a complex, multi-faceted problem is the same as having the answer to the problem.

Check the website again, by the way; this was posted by someone in 2004. The code to do this update has existed since the first errata were ever relesed for the involved processor lines. Microcode in desktop processors is hardly new, and in processors in general it's a mature idea. I don't think it's something that was "released late".

 

[AreEss]

I actually have tried using Registered ECC memory with an A64 by mistake once. An Athlon 64 will not even post with it. The boards can use ECC memory, but not Registered ECC which is different.

In my experience, even with Athlons, it's usually a board issue over a memory controller issue. KT7-RAIDs will run ECC Registered even though the KT133 chipset supposedly doesn't support it. Plus most 'ECC' isn't ECC, it's EC. (e.g. Micron extra-chip stuff.) ECC only happens when you've got the parity checking on the DIMM. Most of the time the BIOS doesn't like the SPD or the voltage which is why it doesn't POST. Just have to find a board that's a little less concerned about how cheap your memory was.

As far as I know, all S939 Opterons are the same as their Athlon 64 and X2 counterparts, reguardless of stepping and revision.

I know that Foxconn's 2200/2050 socket 939 requires ECC Registered. (NFPsomethingsomething. It's like 16 characters long.) Otherwise though, I haven't found any actual Opteron 939 boards worth testing. I'm liking the Foxconn though, and would probably use one for my own workstation at this rate, if it had PCI-X.

 

[draksia]

I know that Foxconn's 2200/2050 socket 939 requires ECC Registered. (NFPsomethingsomething. It's like 16 characters long.) Otherwise though, I haven't found any actual Opteron 939 boards worth testing. I'm liking the Foxconn though, and would probably use one for my own workstation at this rate, if it had PCI-X.

Are you talking about this one ?

http://www.foxconnchannel.com/products_motherboard_2.cfm?pName=NFPIK8AA-8EKRS

Because thats socket940 and takes regular old opterons not the new A64 renames that are socket939. Which does in fact take registered ram because its the 940 operton.

I am not aware of any socket939 boards that use the nForce professional 2200 and 2050. I don't think it is even possible because the socket939 chips lack the hyper transport links to talk to two different chipsets like 940 opterons can.

 

[AreEss]

Are you talking about this one ?

http://www.foxconnchannel.com/products_motherboard_2.cfm?pName=NFPIK8AA-8EKRS

Because thats socket940 and takes regular old opterons not the new A64 renames that are socket939. Which does in fact take registered ram because its the 940 operton.

I am not aware of any socket939 boards that use the nForce professional 2200 and 2050. I don't think it is even possible because the socket939 chips lack the hyper transport links to talk to two different chipsets like 940 opterons can.

Bah, my own fault for trusting a confused vendor. The one I received had the CPU already installed, E4, I presumed 'okay, yeah, 939.' But yeah, pulled it, 940 not 939. Still an impressive board.

As far as the 2200/2050, I've heard a lot of myth and ignorance about them. 'REQUIRES TWO CPUS!!!' Nope. Opteron 1xx's have the same HT count as FX's, which are 939's. The 2200/2050 don't require X amount of HT links; it's preferred, but not required. The actual requirements are determined by how the board itself was designed and laid out.