<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN">
<html>
<head>
<meta content="text/html; charset=ISO-8859-1"
http-equiv="Content-Type">
</head>
<body bgcolor="#ffffff" text="#000000">
On 2014-04-23 21:24, Gordon Dey wrote:
<blockquote cite="mid:535867E1.2050303@happydeys.ca" type="cite">
<pre wrap="">For these older CPU Intel chips, the memory controller architecture
could take advantage of symmetric memory bank pairs. (google this for
more info, there is an Intel white paper on memory architecture.) That
is, 2x2GB sticks performs better than 1x4GB and so on. So I would
advise looking for memory in pairs and in powers of 2 (1, 2, 4, 8, 16)</pre>
</blockquote>
<big>I may be able to help here, if I do not wind up in the weeds with
this description.<br>
<br>
One reason for a symmetric, banked memory was that there is a
requirement for a minimum time between successive reads and writes to
and from semiconductor memory devices. The clock rate of the memory
subsystem could be matched to achieve this requirement, i.e. slowed
down.<br>
<br>
Someone decided to move what the CPU would see as successive addresses
to different physical devices and use hardware to juggle which device
the CPU was looking at.<br>
<br>
For example, I will show this using bytes:<br>
<br>
Address 0x0000: would be in device A<br>
Address 0x0001: would be in device B<br>
Address 0x0002: would be in device A<br>
Address 0x0003: would be in device B<br>
and so on.<br>
<br>
The CPU would see a contiguous memory space from address 0x00, 0x01,
0x02 through to 0xFFFF.<br>
<br>
It is thought that the distribution of data and executable code would
most probably be such that both data and executable code, would be
addressed such that successive accesses to the same physical device
would not be as often as accesses to the alternate devices. If there
were successive accesses to the same device, hardware tricks would come
into play to do things like stretch the clock cycle </big><big>for the
duration of the access, </big><big>thus maintaining the timing
requirement for successive accesses to the same device.<br>
<br>
Of course, 8 bit memory in a PC is a few decades past. But the example
holds for 16, 32 and 64 bit memory.<br>
(And with bytes, I am skipping over a discussion of the way different
tribes eat eggs in Gulliver's Travels.)<br>
<br>
However, with the 32 and 64 bit memory devices, the "interleaving" is
still present. But may be either built in to the chips themselves or
into the memory modules. Or the "interleaving" is built into the chip
sets on the mother board. <br>
<br>
And, if I understand correctly, the symmetric arrangement is built into
the memory modules as opposed to the way motherboard sockets are
populated</big><big> with memory modules. (Ref: Gordon's text above.)</big><br>
<big><br>
As Gordon suggested, or implied, with older hardware, the memory
modules might not have a built-in symmetric arrangement of the physical
memory leaving this up to the memory management subsystems on the
mother board. And </big><big>to the way motherboard sockets are
populated with memory modules by the user.</big><br>
<big><br>
John Johnson<br>
<br>
<br>
</big>
<br>
</body>
</html>