DDR2 Quad-Core Ready AMD Systems
Original Article Date: 2006-10-26
Following the successful release of Intel's Dempsey and Woodcrest Xeon chips this
summer, AMD is hitting back!
A couple of months ago, AMD released their second generation of Opteron CPUs with
the AM2 and Socket F designs, but I wanted to wait until full platform support was
available before going in to what the impact of these new chips entails.
The good news is that the new chips are no more
expensive than the previous generation
of dual-core Opterons. And platforms built using the new sockets can be upgraded
later to quad-core and possibly even octo-core when those chips are released.
Let's start with the basic differences between AM2 and Socket F, and how they are
different from their predecessors.
Changes over the First-Generation Opteron Chips
The most important advance in the new socket designs is that they will be able to
support multiple-core chips, specifically quad-core CPUs that are
expected to be released in the first half of 2007. Single-core chips are not supported,
so ALL chips currently being sold for AM2 and Socket F are dual-core.
This support of multiple core CPUs is the best reason to buy into the new Socket
designs now, as it means that you will be able to upgrade your servers or workstations in, say, one or two years time by adding quad or possibly octo-core CPUs, potentially
doubling or quadrupling system performance
without having to replace anything except
Some users may be concerned about software licensing issues as a result. Microsoft
and the main Linux vendors have committed themselves to licensing per socket
and not per core, so Windows XP will run on anything up to two sockets, regardless
of how many cores are embedded upon each socket. Other software
may not be so generous, especially in the engineering and graphical modeling sectors.
To prevent any license violations, however, you can disable the second core on each
CPU in BIOS, so that you're effectively running only single-core again. Whilst this
means you're only getting half the power you can potentially run, your license will
not be compromised if you have only single-CPU licenses.
The other significant change is the switch to DDR2 RAM. The previous
generation had been using DDR RAM, the standard of which is now almost a decade
The clock speed on DDR was limited to 200MHz (400MHz double data rate), so to get
higher RAM clock speeds running Opteron chips, it was necessary to move to DDR2.
The AM2 socket will currently support up to 400MHz RAM clock speed (800MHz
double data rate), whilst Socket F will support up to 333MHz (667MHz
double data rate). To an extent, this speed increase is offset by longer CAS Latencies
of 4 or 5 clock cycles compared to 2.5 or 3, but the overall result is higher bandwidth
and faster fetching of data.
DDR was never really expected to be accommodating 1 or 2GB of RAM on each DIMM,
and so for servers and workstations, which routinely require such high RAM densities,
scalability became an issue, and as a consequence these DIMMs became expensive.
The DDR2 standard was designed to allow for these levels of memory density, and
so in time you can expect DDR2 DIMMs to become much cheaper per GB than DDR.
DDR2 is also more efficient on power than DDR, and will result in less cooling-related
issues in servers, and lower electricity
AM2 vs Socket F
AMD have made a distinction between their two new CPU form factors primarily for
AM2 is primarily for the desktop market, and is exclusively
single socket whilst Socket F is aimed at servers
and workstations, allowing 2-way, 4-way and 8-way socket solutions.
AM2 uses a 940-connector Land Grid Array, or LGA (i.e. pinless
CPU design - the pins are on the mainboard "socket"), whilst Socket F uses a
1207-connector LGA. An extra 247 pins are required on the Socket F
design to accomodate extra RAM bandwidth (F supports up to 8 DIMMs per socket, whilst
AM2 allows only 4). Inter-socket Hyper-Transport links on the Socket F design, which
allow more than one
CPU socket to work together also require extra bandwidth, and
more pins as a consequence.
The most obvious similarities between the two socket designs are that they're currently
both dual-core (and will be ready to handle quad-core when those chips are released,
expected in Q1 or Q2 of 2007), and both now use DDR2 RAM. In the case of AM2 however,
the supported RAM is unbuffered, whilst the Socket F requires ECC Registered RAM.
L2 cache size is also the same in this current generation of chips - 1MB per core.
As they are currently dual-core chips, this therefore means 2MB per socket, for
both AM2 and Socket F.
Deciphering the New Numbering System
I know... despite the best efforts of AMD to get us into processor "numbering" we're
still fixated on clock speed. One day, clock speed won't be the
headline performance indicator that we look for. And rightly so. Clock speed isn't
the be-all and end-all that it once was. With dual-core and soon-to-be quad core
chips available, this "horizontal" movement outwards has already become more significant
in terms of real-world server and workstation performance than the "vertical" clock
speed of yesterday's CPUs. And remember that with cooling issues and electricity
bills affecting overall TCO, higher clock speed is often a bad thing, as those higher
clock speed models usually have higher wattage values.
So AMD were the first to introduce a performance numbering system
with their first generation Opteron CPUs (e.g. 146, 250, 880 etc.). Intel followed
suit, when they realised they weren't going to make a 4GHz clock speed chip a reality,
and started having to look at multi-core to get extra performance. So get used it.
Processor performance numbering is here to stay!
In order to decipher the new Opteron numbering system, there are only two things
you really need to look at in the four number code they have.
The first number is the same as the old three digit Opteron numbers,
namely, how many ways you can configure the CPU in a system. 1XXX Opterons can only
be run in a single socket, 2XXX Opterons can be run in up to two sockets, and 8XXX
Opterons can be run in up to eight socket configurations.
The second number is always a 2. This just means that the chips are Opteron generation
number 2, so isn't something you need to focus on.
The third and fourth numbers are the indicator of performance.
A higher number means better performance. As all the chips are dual-core now, this
performance increase is simply determined by clock speed, e.g. a 2210 Opteron runs
at 1.8GHz, whilst a 2220 Opteron runs at 2.8GHz. When quad-core chips are released
next year, however, expect even higher third and fourth numbers.
AMD have produced a great web page explaining this
New System Line-Up
We have overhauled our line-up of 1-way and 2-way servers and workstations.
Our 4-way and 8-way servers are still built on the tried and tested Socket 940 design,
as our preferred Tyan platform support for the 8200 series DDR2 8-way capable Opterons
is still not yet available (expected late Q4 06, early Q1 07).
Performance and price are about the same as the current generation of Socket 939/940
Opterons - so there isn't much of a change there. The great advantage about the
new platforms is simply that they will be upgradeable to future multi-core CPUs
as and when they are released. So with a simple CPU swap-out, you'll be able to
double or quadruple your server in no time at all. These systems are future-proofed!
Chief Systems Engineer