Quoting Colin Fee (tfeccles@gmail.com):
AIUI it's an economic compromise made by CPU manufacturers and mobo manufacturers.
Sounds reasonable (and this is what I expected). Part of what I'm curious about (not asking _you_ specifically, but rather just putting the question out in the air) is whether this implementation compromise is, particularly in recently discussed SoCs and motherboards, imposed by the SoC or not. And which specific component or components create(s) it. Most particularly, I am curious whether I am wasting my time looking aroudn for motherboards compatible with AMD's 2014 'Puma' series of SoCs (the Beema and Mullins series) capable of supporting more than 8GB of RAM -- because to the extent I find Beema/Mullins used, which is distressingly rare, I see 8GB limits. I am vacillating on the brink of buying a SoC / motherboard combo based on the 2013 chips taht the 2014 'Puma' architecture directly replaced, namely the 'Jaguar' architecture 'Kabini' SoCs, specifically because the older 2013 SoC maxes out at 32GB supported RAM rather than the newer, 2014 replacement SoC's (seeming?) 8GB limit. The point is, I don't want to miss better, one-year-newer, hardware out of ignorance -- if higher-RAM 'Puma' motherboards do exist. In addition, the (seeming) 4x decline in RAM capacity with the successor 2014 chip series seems awfully strange, and I'm trying to understand what happened. (FWIW, I was an EECS aka Electrical Engineering and Computer Science -- pronounced 'eeks!' -- major at a certain Ivy League university named for William Prince of Nassau, back in unfathomably ancient days, not that that does me any good in this century. I switched to maths, which Yanks call 'math', before graduation. Ridiculoously long ago, and, like essentially all teenagers, I had no clue whatsoever what I was doing.)
Whilst 64 bit CPUs have a theoretical 64 bit address space available to them for performance and other functionality reasons the CPU designers limit the number address lines i.e pins available to memory. For example the Kabini CPUs featured in this thread have 40-bit address space, whereas some the higher end CPUs will have 42 and 43 bit address lines available to them.
Doesn't 40-bit address space (2^40) equate to 1 terabyte? All the Kabini-compatible motherboards I'd seen max out at 32GB, which is something like 2^36. So, if you say Kabini has 40-bit address space, I'm not sure where that has any real-world manifestation. I often make a fencepost error in these shirtsleeve calculations, and real hardware engineers keep track of 'Well, it's actually 2^36 - 1' sorts of edge conditions, but I'm sure I'm no more than off by a factor of two, here. FWIW, if in 2015, actually _had_ ability to support 1 TB-ish RAM (within a factor of two, 2^40 addresses) on Kabini or anything similar, I'd feel extremely future-proofed indeed, and consider the decade of the 2030 able to take care of itself. I've not seen that at all -- much less on those chips' next-year replacements.
Of course this is larger than some of the 32Gb limits we see on mobos. Here we hit the economic compromises made by the mobo makers.
This is what I figure -- and is very small misfortune compared to death and taxes. But I remain curious about _how_ (where) that gets implemented, and why AMD (apparently) decreed a further 4x cutback in RAM capacity when it replaced Kabini/Temash with its direct successor.
I'm sure if you looked around you could find a mobo that maxed out the RAM limit but I expect it would be expensive, power hungry and large (well full-sized anyway).
Well, yes. In particular, some of the HP ProLiant rackmount colo boxes I've recently worked with professionally can hold very high amounts of RAM -- and I'm glad I don't pay the power bill for them (or hear their banshee screaming 24x7 in my house). -- Cheers, I'm ashamed at how often I use a thesaurus. I mean bashful. Rick Moen Embarrassed! Wait--humiliated. Repentant. Chagrined! Sh*t! rick@linuxmafia.com -- @cinemasins McQ! (4x80)