Remembrance of things future
by Alan Zisman
(c) 1997. First
published in Canadian Computer Wholesaler, May 1997
In this series, we?re looking at changes in the
hardware?how new features
in PC-design promise to continue to provide users with machines that
will
do more?not only faster but easier. We?ve looked at Intel?s new family
of MMX processors, now available, and at Universal Serial Bus,
Firewire,
and the Accelerated Graphics Port?all somewhere in between promise and
reality.
But none of these advances will work without
memory?the chips the computer
uses to remember what it?s just done and what it?s about to do.
Many users still get confused between storage?their
hard drive, for
example, and memory?the computer?s RAM. Perhaps because both can be
measured
in Megabytes. The difference is simple?imagine looking up a recipe in a
cookbook, and then trying to cook it. The cookbook is like your hard
drive?information
in long-term storage. But in order to make dinner, you need to read the
recipe, and remember it?at least long enough to cook the meal. Your
brain,
in this case, is acting like the computer?s RAM? remembering just
enough
of the recipe to get you from the cookbook to the stove.
And more ram is always better. In our recipe example,
more RAM would
mean being able to remember several of the cookbook?s instructions at
once,
instead of having to go back and read each step, one at a time. As a
result,
adding more ram makes any computer run faster and more efficiently.
But while our computers sport ever-faster CPUs, the
speed of the RAM
hasn?t kept up. The original IBM PC featured an Intel 8088 processor,
running
at a blazing 4.77 mhz, and came standard with between 16 and 64 kb of
200
nanosecond (nsec) RAM. Your 200 mhz Pentium processor is 40 times
faster,
and about 400 times as powerful as that original processor. And with 16
megs or more RAM, it?s got 1,000 times as much memory. But that memory
is probably running at 60 nsec?not even four times as fast as that on
the
1981 PC. Designers have had to build in a series of tricks to keep the
processor and the RAM working together:
? Wait States. Standard DRAM (Dynamic RAM) needs to be
refreshed?electrically
recharged with its data. Designers add ?wait states??periods where the
CPU sits idle, waiting for the DRAM to be refreshed. Faster RAM means
fewer
wait states.
? RAM caches. Most often, CPUs access the same instructions over and
over again. (?Peel a carrot. Slice it. Place in frying pan. Stir. Peel
a carrot??) As a result, a little bit of very fast RAM can go a long
way
towards meeting the CPU?s needs, letting the CPU access the slower main
RAM only when something isn?t in the cache. Typically, modern computers
have several levels of cache-ram. A tiny bit of very fast ram is built
right into the CPU itself (much of the performance increase of the new
MMX CPUs comes from doubling the size of this on-board cache). Next,
there?s
the so-called Level 2 (L2) cache on the motherboard? typically
256-512kb
of fast but expensive SRAM (Static RAM, which doesn?t need to be
constantly
refreshed). Much of the power of the Pentium-Pro design comes from it
including
the L-2 cache right on the chip.
? Newer and faster RAM designs. Over the past couple of years, old
standard DRAM was replaced with more efficient Fast Page Mode (FPM)
RAM.
That in turn was replaced with today?s standard?Extended Data Out
(EDO)
RAM. By letting the CPU read and write at the same time as the refresh
cycle, this speeds up access and eliminates wait states. But even EDO
is
reaching its limits, as processors get faster and faster.
Replacing EDO is Synchronous DRAM (SDRAM); because it
can run in higher
speed systems, it?s poised to become the new standard for 1997-98. But
increases in processor speed that are already in the works limit the
potential
of SDRAM much beyond that.
Coming right at us is Rambus DRAM (RDRAM).
Rambus is a hot company,
with a recent, wildly successful stock offering, based on owning the
next
generation of RAM technology. Early versions of its RAM are available
in
high-end Silicon Graphics workstations, and in $300 Nintendo-64 game
machines,
as well as on some PC graphics cards.
While more standard RAM communicates with the CPU in
16 or 32 bit pieces,
Rambus-RAM uses a humble 8-bit interface. It overcomes this
seeming-limitation
because the interface is extremely fast?currently at 250 MHz, speeds of
600 MHz have been demonstrated, with even faster versions on the
horizon.
To make use of this super-fast memory design,
motherboard companies
will need to license the custom Rambus Channel?the special chip set
that
allows the CPU to interact with the RAM. CPU and motherboard giant
Intel
has thrown its support behind Rambus?the two companies have announced
plans
to cooperate to design the next generation, being called nDRAM. They?re
hoping to reach speeds up to 1.6 GBps (Gigabytes per second!) by 1999,
just in time for Intel?s 64-bit P7 Merced processor.
Until that becomes reality, there are other shifts in
how RAM is being
sold. 386 and many 486 generation computers typically used so-called
30-pin
SIMMs. These were 8-bit pieces (ironically, like the futuristing
RDRAM),
and had to be added in pairs for 16-bit 386SX busses, and in sets of
four
for 32-bit 386DX and 486 systems.
Pentium systems, however, use a 64-bit memory bus? to
use 30-pin SIMMs,
users would have to install these eight at a time, an impractical
arrangement.
So virtually all such systems were built to use newer, 32-bit 72-pin
SIMMs.
These could be used singly in a 486, or in matched pairs on a Pentium.
New Pentium and Pentium-Pro motherboards have, however, stolen an idea
from recent Macintosh designs, and now include slots for 64-bit DIMMs
(dual
in-line memory modules). A single DIMM can replace a pair of SIMMs?an
advantage
as users try and cram more and more memory into their case.
But that?s not the only Mac design showing up on PCs
in the near future.
Currently, PCs end up with RAM in several places.
There?s the main system
RAM, of course. But there?s also RAM on the video card?typically 2 ? 4
megs worth. There may even be RAM on a wavetable sound card, for
storing
sound samples, cache ram on a high-end disk controller, and other bits
and pieces around the system. For several years, Macintoshes have been
designed using a Unified Memory Architecture (UMA)? a simpler design in
which the main system RAM is shared around as needed.
When RAM was expensive, this made sense? you might
have 4 meg of RAM
on your video card, but only be using 1 meg or so, depending on the
number
of colours and screen resolution?the rest of the video RAM would be
going
to waste. Implementing UMA could, as a result, lower system prices.
UMA, however, has been a mixed benefit. Performance of
UMA systems has
tended to be lower than systems with dedicated graphics RAM; this is
especially
true when running at high video resolution and colour depth, which
requires
a large amount of RAM?under UMA, this deprives the CPU access to a
significant
amount of RAM. As well, constantly deciding how to share the system RAM
around is just one more task to dump onto the CPU.
Operating systems have to be designed to support
UMA?this support is
not currently written into either Windows 95 or NT, but is expected to
show up in future generations.
That?s because the performance hit associated with UMA
will become less
of an issue in the future?with more powerful CPUs accessing large pools
of system RAM. As a result, expect to see this becoming a standard
feature
on low-mid-range systems over the next few years.
For now, make sure your systems support DIMMs; these
will be needed
to support the large amounts of RAM users will be requiring. EDO RAM is
on the way out?SDRAM will be the standard by late 1997, through 1998.
After
that, look for the results of the Intel-Rambus cooperation to take us
through
the turn of the century.
Avoid Unified Memory Architecture models for now?but
expect this to
become common on entry-level systems in a year or so.
Now if only I could get this computer to make dinner
for me!
Next month?making PCs simpler: the NetPC and the
Zero-Administration
initiative.