Opteron - single dual core vs two single cores

  • Thread starter Thread starter CharlesBlackstone
  • Start date Start date
C

CharlesBlackstone

Hi, I'm about to buy a workstation, and wonder what the functional
difference would be between getting a machine with two single-core
processors, and a machine with one dual-core processor? Is it
essentially the same thing, or are the two single-cores more
independent in terms of function, memory access, etc. that would affect
overall performance?

Thanks very much......
 
CharlesBlackstone said:
Hi, I'm about to buy a workstation, and wonder what the functional
difference would be between getting a machine with two single-core
processors, and a machine with one dual-core processor? Is it
essentially the same thing, or are the two single-cores more
independent in terms of function, memory access, etc. that would affect
overall performance?

Thanks very much......
Click to expand...

Two single-cores versus a dual core:
http://www.geek.com/news/geeknews/2005Nov/bch20051108033166.htm

Benchmarks start on page 10:
http://www.tomshardware.com/2005/11/07/single/page10.html

The advantage of two single core processors, is the number
of chipset chips on the motherboard. The AMD processor is
a "hub" with Hypertransport busses sprouting out of it.
With two single cores, you can have two true PCI Express X16
slots, plus lots of tunnels for peripheral busses. If
there is only one CPU socket on a motherboard, generally
the I/O features are a bit less (they don't have to be,
but companies don't generally use large numbers of daisy
chained tunnel chips). This is mostly related to motherboard
marketing, and the proposed uses for single socket versus
dual socket systems (workstation versus server perhaps).

Example block diagram on page 2, of a dual socket board:
ftp://ftp.tyan.com/datasheets/d_s2895_100.pdf

Note that if the "H0" socket is not occupied by a processor,
you lose the use of a PCI Express x16 slot, an ethernet
interface, and the use of four more DIMM slots. The board
will still work with just one processor (single or dual core)
in socket "H1". If you buy a motherboard with a single socket,
well, you never have to worry about this, but you also have
fewer facilities of all sorts on the motherboard.

If your box is a pure compute engine, and the only peripheral
you have is a USB socket for your Ipod, the single socket
dual core processor wins. If the box is intended to hold
a wealth of hardware, then using two sockets might be a
better answer, even if it is computationally slightly slower.

With two sockets, you have two memory controllers, and
you would think, more memory bandwidth. A single socket
setup has access to fewer total sticks of RAM, which
might be important consideration if the problems you
handle need a lot of RAM.

"Cache-Coherent Nonuniform Memory Access (ccNUMA)" - the theory
http://cdrom.amd.com/21860/updates/...ext=OptguideHelp&file=OptguideHelp-09-04.html

"Non-Uniform Memory Architecture (NUMA): Dual Processor
AMD Opteron Platform Analysis" - benchmarks
(Conclusion - set node interleave off ; use "numa aware" OS)
http://www.digit-life.com/articles2/cpu/rmma-numa.html

That last article shows that you can get 12.8GB/sec memory
bandwidth, as long as each single core processor only accesses
its own local memory at 6.4GB/sec. While the dual socket
system has the ability to use any memory by either processor,
there is a penalty associated with going to the "other" memory.

If your system has ECC memory, I think features like scrubbing
and the like, may also affect whether things like node
interleave can be turned on.

Dual sockets have their complexities, as do setting up dual core
processors. For the chunk of cash you'll be paying for either
system, I hope the person setting it up does a good job.

Paul
 
Hi, I'm about to buy a workstation, and wonder what the functional
difference would be between getting a machine with two single-core
processors, and a machine with one dual-core processor? Is it
essentially the same thing, or are the two single-cores more
independent in terms of function, memory access, etc. that would affect
overall performance?
Click to expand...

For workstations, 1 core per socket is kinda dead. Go for the dual-core.
The X2's just came down, a lot.

Besides, I don't believe AMD is selling 940-pin single core Opterons
anymore, anyway.
 
Paul said:
Two single-cores versus a dual core:
http://www.geek.com/news/geeknews/2005Nov/bch20051108033166.htm

Benchmarks start on page 10:
http://www.tomshardware.com/2005/11/07/single/page10.html

The advantage of two single core processors, is the number
of chipset chips on the motherboard. The AMD processor is
a "hub" with Hypertransport busses sprouting out of it.
With two single cores, you can have two true PCI Express X16
slots, plus lots of tunnels for peripheral busses. If
there is only one CPU socket on a motherboard, generally
the I/O features are a bit less (they don't have to be,
but companies don't generally use large numbers of daisy
chained tunnel chips). This is mostly related to motherboard
marketing, and the proposed uses for single socket versus
dual socket systems (workstation versus server perhaps).

Example block diagram on page 2, of a dual socket board:
ftp://ftp.tyan.com/datasheets/d_s2895_100.pdf

Note that if the "H0" socket is not occupied by a processor,
you lose the use of a PCI Express x16 slot, an ethernet
interface, and the use of four more DIMM slots. The board
will still work with just one processor (single or dual core)
in socket "H1". If you buy a motherboard with a single socket,
well, you never have to worry about this, but you also have
fewer facilities of all sorts on the motherboard.

If your box is a pure compute engine, and the only peripheral
you have is a USB socket for your Ipod, the single socket
dual core processor wins. If the box is intended to hold
a wealth of hardware, then using two sockets might be a
better answer, even if it is computationally slightly slower.

With two sockets, you have two memory controllers, and
you would think, more memory bandwidth. A single socket
setup has access to fewer total sticks of RAM, which
might be important consideration if the problems you
handle need a lot of RAM.

"Cache-Coherent Nonuniform Memory Access (ccNUMA)" - the theory
http://cdrom.amd.com/21860/updates/...ext=OptguideHelp&file=OptguideHelp-09-04.html

"Non-Uniform Memory Architecture (NUMA): Dual Processor
AMD Opteron Platform Analysis" - benchmarks
(Conclusion - set node interleave off ; use "numa aware" OS)
http://www.digit-life.com/articles2/cpu/rmma-numa.html

That last article shows that you can get 12.8GB/sec memory
bandwidth, as long as each single core processor only accesses
its own local memory at 6.4GB/sec. While the dual socket
system has the ability to use any memory by either processor,
there is a penalty associated with going to the "other" memory.

If your system has ECC memory, I think features like scrubbing
and the like, may also affect whether things like node
interleave can be turned on.

Dual sockets have their complexities, as do setting up dual core
processors. For the chunk of cash you'll be paying for either
system, I hope the person setting it up does a good job.

Paul
Click to expand...


Paul, many, many thanks for your excellent answer. I sure appreciate
it...

JK
 
Paul said:
Two single-cores versus a dual core:
http://www.geek.com/news/geeknews/2005Nov/bch20051108033166.htm

Benchmarks start on page 10:
http://www.tomshardware.com/2005/11/07/single/page10.html

The advantage of two single core processors, is the number
of chipset chips on the motherboard. The AMD processor is
a "hub" with Hypertransport busses sprouting out of it.
With two single cores, you can have two true PCI Express X16
slots, plus lots of tunnels for peripheral busses. If
there is only one CPU socket on a motherboard, generally
the I/O features are a bit less (they don't have to be,
but companies don't generally use large numbers of daisy
chained tunnel chips). This is mostly related to motherboard
marketing, and the proposed uses for single socket versus
dual socket systems (workstation versus server perhaps).

Example block diagram on page 2, of a dual socket board:
ftp://ftp.tyan.com/datasheets/d_s2895_100.pdf

Note that if the "H0" socket is not occupied by a processor,
you lose the use of a PCI Express x16 slot, an ethernet
interface, and the use of four more DIMM slots. The board
will still work with just one processor (single or dual core)
in socket "H1". If you buy a motherboard with a single socket,
well, you never have to worry about this, but you also have
fewer facilities of all sorts on the motherboard.

If your box is a pure compute engine, and the only peripheral
you have is a USB socket for your Ipod, the single socket
dual core processor wins. If the box is intended to hold
a wealth of hardware, then using two sockets might be a
better answer, even if it is computationally slightly slower.

With two sockets, you have two memory controllers, and
you would think, more memory bandwidth. A single socket
setup has access to fewer total sticks of RAM, which
might be important consideration if the problems you
handle need a lot of RAM.

"Cache-Coherent Nonuniform Memory Access (ccNUMA)" - the theory
http://cdrom.amd.com/21860/updates/...ext=OptguideHelp&file=OptguideHelp-09-04.html

"Non-Uniform Memory Architecture (NUMA): Dual Processor
AMD Opteron Platform Analysis" - benchmarks
(Conclusion - set node interleave off ; use "numa aware" OS)
http://www.digit-life.com/articles2/cpu/rmma-numa.html

That last article shows that you can get 12.8GB/sec memory
bandwidth, as long as each single core processor only accesses
its own local memory at 6.4GB/sec. While the dual socket
system has the ability to use any memory by either processor,
there is a penalty associated with going to the "other" memory.

If your system has ECC memory, I think features like scrubbing
and the like, may also affect whether things like node
interleave can be turned on.

Dual sockets have their complexities, as do setting up dual core
processors. For the chunk of cash you'll be paying for either
system, I hope the person setting it up does a good job.

Paul
Click to expand...


Paul, many, many thanks for your excellent answer. I sure appreciate
it...

JK
 
Hi, I'm about to buy a workstation, and wonder what the functional
difference would be between getting a machine with two single-core
processors, and a machine with one dual-core processor? Is it
essentially the same thing, or are the two single-cores more
independent in terms of function, memory access, etc. that would affect
overall performance?
Click to expand...

There are some differences between the two, the biggest being that
with a single dual-core processor you have all your memory hanging off
one memory controller while a pair of single-core chips each have
memory hanging off their own memory controllers. The
advantage/disadvantage here is that with the one dual-core chip you
can run with only 2 DIMMs but are limited to a maximum of 4 or 8 DIMMs
(depending on the board), while with the two single-core chips you
want a minimum of 4 DIMMs but can expand up to 8 or 16 DIMMs.
Therefore, if you're application needs lots of memory (8GB+) a pair of
single-core chips is processors might be better.

Another difference is that the single-core chips usually have one
clock speed above the fastest dual-core chips. For example, right now
the fastest dual-core Opteron is the 285, which runs at 2.6GHz. The
fastest single-core Opteron is the 254 which runs at 2.8GHz.

In terms of performance, there are a handful of pluses and minuses on
either side. Two processors, each with their own memory controller,
give you more theoretical bandwidth. The downside is that you now
have a NUMA setup with half your memory being remote and therefore
suffering a latency penalty. Cache access also can suffer a bit of a
latency penalty as accessing the cache on a second core in a dual-core
chip is much faster then accessing cache on a physically different
chip in a dual-processor setup.

In the end though, the performance differences mostly get lost in the
noise. A single dual-core chip will usually come out on top, but not
by any big margin. The *REAL* difference in all this is cost. A
one-socket motherboard and a single dual-core chip will be a fair bit
cheaper than a 2P motherboard with a pair of single-core chips. This
is especially true if you opt for a dual-core Athlon64 X2 vs. a pair
of single-core Opteron chips. Even though the X2 chips are marketed
as desktop chips they could make for quite a good workstation setup
with the right motherboard. Only possible shortfall here is that
unbuffered ECC memory is a bit rare, so if error correction is
required for your workstation you might need to do a bit of digging.
 
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

There are some differences between the two, the biggest being that
with a single dual-core processor you have all your memory hanging off
one memory controller while a pair of single-core chips each have
memory hanging off their own memory controllers.
Click to expand...

OTOH, the two cores in a dual-core processor have equal access to memory.
In a dual-processor system with memory attached to one processor, processor
#2 will have to go through processor #1 to access memory. The dual-Opteron
at which I'm typing this message is set up that way, with a couple of 512-MB
DIMMs hanging off one processor. Prime95 running on processor #2 takes
noticeably longer per iteration than Prime95 running on processor #1. (I'd
provide numbers, but they're currently doing different types of work (one is
factoring and the other is LL-testing).)

Putting memory on both processors would fix this, but that would be an
additional expense you would want to keep in mind.

_/_
/ v \ Scott Alfter (remove the obvious to send mail)
(IIGS( http://alfter.us/ Top-posting!
\_^_/ rm -rf /bin/laden >What's the most annoying thing on Usenet?

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.5 (GNU/Linux)

iD8DBQFE2icwVgTKos01OwkRAtJcAKDggnrlJqP9Pur4R+Je3SNPVAQC+gCgvAD5
nRYA8cUy2xoOc+dXdrZ7v9I=
=xCHb
-----END PGP SIGNATURE-----
 
There are some differences between the two, the biggest being that
with a single dual-core processor you have all your memory hanging off
one memory controller while a pair of single-core chips each have
memory hanging off their own memory controllers. The
advantage/disadvantage here is that with the one dual-core chip you
can run with only 2 DIMMs but are limited to a maximum of 4 or 8 DIMMs
(depending on the board), while with the two single-core chips you
want a minimum of 4 DIMMs but can expand up to 8 or 16 DIMMs.
Therefore, if you're application needs lots of memory (8GB+) a pair of
single-core chips is processors might be better.

Another difference is that the single-core chips usually have one
clock speed above the fastest dual-core chips. For example, right now
the fastest dual-core Opteron is the 285, which runs at 2.6GHz. The
fastest single-core Opteron is the 254 which runs at 2.8GHz.

In terms of performance, there are a handful of pluses and minuses on
either side. Two processors, each with their own memory controller,
give you more theoretical bandwidth. The downside is that you now
have a NUMA setup with half your memory being remote and therefore
suffering a latency penalty. Cache access also can suffer a bit of a
latency penalty as accessing the cache on a second core in a dual-core
chip is much faster then accessing cache on a physically different
chip in a dual-processor setup.

In the end though, the performance differences mostly get lost in the
noise. A single dual-core chip will usually come out on top, but not
by any big margin. The *REAL* difference in all this is cost. A
one-socket motherboard and a single dual-core chip will be a fair bit
cheaper than a 2P motherboard with a pair of single-core chips. This
is especially true if you opt for a dual-core Athlon64 X2 vs. a pair
of single-core Opteron chips. Even though the X2 chips are marketed
as desktop chips they could make for quite a good workstation setup
with the right motherboard. Only possible shortfall here is that
unbuffered ECC memory is a bit rare, so if error correction is
required for your workstation you might need to do a bit of digging.
Click to expand...

I didn't have any trouble getting unbuffered ECC DDR a few weeks ago and
yes a system like that makes a nice workstation or Intranet Web/ASP server
-- my purpose -- for not a lot of money. Funny, Crucial doesn't list the
2GB kit I got in their DDR listings but looking up the mbrd, an Asus A8N-E,
still shows it as available; for DDR2, they show it in the PC-5300 stuff
*but* at 3x the price of the DDR!!... which makes s939 still an attractive
bet, to me anyway. Of course you won't find it in the "boutique" lines
which cater to the OC/gamer market and which many on-line vendors also seem
to concentrate on.

Oh and BTW I can find nothing on ECC reporting so who knows if it really
works... whats new??:-)
 
[...]
In the end though, the performance differences mostly get lost in the
noise. A single dual-core chip will usually come out on top, but not
by any big margin. The *REAL* difference in all this is cost. A
one-socket motherboard and a single dual-core chip will be a fair bit
cheaper than a 2P motherboard with a pair of single-core chips.
Click to expand...

You also need to factor in commercial OS licencing. MS views dual-core as a
single CPU for licencing purposes. You pay more to licence Windows on 2
CPUs.
 
* DRS:
You also need to factor in commercial OS licencing. MS views dual-core as a
single CPU for licencing purposes. You pay more to licence Windows on 2
CPUs.
Click to expand...

Nope. Windowsxp Pro works fine with up to two cpus no matter how much
cores it has...

Benjamin
 
Benjamin Gawert said:
* DRS:


Nope. Windowsxp Pro works fine with up to two cpus no matter how much
cores it has...
Click to expand...

My bad, multiprocessor licencing is currently restricted to servers.
 
DRS said:
You also need to factor in commercial OS licencing. MS views dual-core as a
single CPU for licencing purposes. You pay more to licence Windows on 2
CPUs.
Click to expand...

Not for OS itself unless you outgrow the low end Windows Server version
- at 4 sockets I think? But you are right about other licenses eg the
per CPU SQL Server licenses.
 
Scott said:
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1



OTOH, the two cores in a dual-core processor have equal access to memory.
In a dual-processor system with memory attached to one processor, processor
#2 will have to go through processor #1 to access memory. The dual-Opteron
at which I'm typing this message is set up that way, with a couple of 512-MB
DIMMs hanging off one processor. Prime95 running on processor #2 takes
noticeably longer per iteration than Prime95 running on processor #1. (I'd
provide numbers, but they're currently doing different types of work (one is
factoring and the other is LL-testing).)

Putting memory on both processors would fix this, but that would be an
additional expense you would want to keep in mind.
Click to expand...

Scott, how do you tell one processor to do one job, and the other
processor to do the other job?

Thanks....
 
Scott, how do you tell one processor to do one job, and the other
processor to do the other job?
Click to expand...

Scott could write a multi-threaded application which uses multiple
threads. Each thread can potentially run on a discrete physical or
logical processing unit.
 
Scott, how do you tell one processor to do one job, and the other
processor to do the other job?
Click to expand...

In WinXP: Task Manager => Processes => Set Affinity (I think, I'm
on a UP Laptop, right now).
 
In WinXP: Task Manager => Processes => Set Affinity (I think, I'm
on a UP Laptop, right now).
Click to expand...

Small correction: TaskManager => Processes => <RMB> on process =>
Set Affinity => select processors process is allowed to run on
 
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Small correction: TaskManager => Processes => <RMB> on process =>
Set Affinity => select processors process is allowed to run on
Click to expand...

With Prime95, there's a menu option (under Advanced) that sets affinity.
Since I have it set up to run as a service, this is more convenient than
having to set affinity in the Task Manager on every reboot.

_/_
/ v \ Scott Alfter (remove the obvious to send mail)
(IIGS( http://alfter.us/ Top-posting!
\_^_/ rm -rf /bin/laden >What's the most annoying thing on Usenet?

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.5 (GNU/Linux)

iD8DBQFE52T4VgTKos01OwkRAlhhAJsE5qj8hJovJGjMByCeuuEaOfsBBwCePKFT
/XhxmvGNsZbHeczET2k01Ac=
=I/Ss
-----END PGP SIGNATURE-----
 
Back
Top