AMD64 and PCI-Express

[JK]

Why do you need PCI Express?

 

[Synapse Syndrome]

Are there any motherboards that combine AMD64 and PCI-Express yet?

Or Intel the only way to go? Don't really want to lose Hyper-Threading, but
the heat and price of Prescott chips is putting me off.

ss.

 

[Synapse Syndrome]

Why do you need PCI Express?

Because I want to buy a new expensive OpenGL graphics card for CAD within a
year's time. It would seem a waste to buy an expensive card that has an AGP
interface when that standard will start to die.

I also like what I have read about how audio streams are all handled
seperately. I know Longhorn is far away, but at least I could make use of
these abilities when it arrives.

Oh, and there's the bandwidth.. Surely I can make use of it when doing some
rendering on twin monitors with music playing and RAID-0 drives churning,
while recording TV, as I do.

If getting a PCI-express motherboard will not cost a lot more I should at
least look into it. I plan to get one with onboard graphics for the time
being. Maybe I'll decide to buy a convention PC, who knows?

ss.

 

[JK]

Because I want to buy a new expensive OpenGL graphics card for CAD within a
year's time. It would seem a waste to buy an expensive card that has an AGP
interface when that standard will start to die.

Why? When the standard dies(perhaps 3+ years?) it will be time for
a new video card, motherboard, and processor.. It is a better idea to
avoid very new technology until the major kinks are ironed out(aside
from the higher expense for the novelty factor). Think about PCI express
for the next motherboard and video card 3 years from now.

 

[JK]

The Athlon 64 is about great 32 bit performance. The ability to run
64 bit code is an option that is basically thrown in for free due to
the low prices and great 32 bit performance the chips have.

 

[Synapse Syndrome]

Why do you need PCI Express?

I also think that starting to buy futureproof PCIx peripheral cards and
making use of the bandwidth immediately is a lot more sensible than buying a
AMD64 just because oyu might start using a 64-bit OS one day.

ss.

 

[Synapse Syndrome]

Why? When the standard dies(perhaps 3+ years?) it will be time for
a new video card, motherboard, and processor.. It is a better idea to
avoid very new technology until the major kinks are ironed out(aside
from the higher expense for the novelty factor). Think about PCI express
for the next motherboard and video card 3 years from now.

Yeah, I've been reading some reviews of PCIx boards, and the onboard
graphics seems to weak for me to make do with, so I'm thinking on getting
the graphics card sooner rather than later.

So the only advantage of PCIx that I can make use of right now would be the
higher general bandwidth, and if I find out what that motherboard that has
the AGP slot I might still go that way. If the board works well I see no
reason not to get it, as it'll have that High Definition Audio onboard and
still have regular PCI slots for RAID, TV-card etc...

That's right isn't it? Even without any actual PCIx peripherals I could
still benefit from the increased bandwidth?

ss.

 

[Synapse Syndrome]

The Athlon 64 is about great 32 bit performance. The ability to run
64 bit code is an option that is basically thrown in for free due to
the low prices and great 32 bit performance the chips have.

Yeah I know. The only thing making me hesitant about an AMD64 is the lack
of Hyper-Threading. I tend to do a lot of rendering which ties up the
computer.

ss.

 

[General Schvantzkoph]

Are there any motherboards that combine AMD64 and PCI-Express yet?

Or Intel the only way to go? Don't really want to lose Hyper-Threading, but
the heat and price of Prescott chips is putting me off.

ss.

The AMD64 runs rings around the P4. I've been running benchmarks on my
Athlon 64 3400+ laptop. In CPU bound applications it's twice as fast as my
2.66GHz Xeon. If there is a lot of IO then it falls to 40% faster but
that's due to the laptop's slow disk (4400RPM vs 7200RPM on the Xeon). To
put that in MHz terms the 2.2GHz Athlon 64 is equivalent to a 4-5GHz
Xeon depending on the application.

For desktop applications the 939pin A64s have twice the memory bandwidth
of the 754 pin A64 in my laptop which means that they should be even
faster. If you get a motherboard that uses the Nvidia NForce 3-250GB or
Ultra bridge chips you'll get an extremely high performance gigabit
ethernet MAC that's connected directly to the hypertransport bus so it's
not stealing any bandwidth from the PCI slots. The SATA controllers are
also built in so they are stealing any PCI bandwidth either. To top it off
the hypertransport bus that connects the CPU to the bridge chip runs
4GBytes/second and it's solely used for IO. Intel chips access their
memory systems over the same frontside bus as their IO systems, the AMDs
have dedicated on chip memory controllers which are not only higher
performance than Intel's (because the latency is much lower) but also
because they aren't competing for bandwidth with the IO systems.

To sum up,

1) Athlon 64 is much faster than the P4
2) Athlon 64 has more bandwidth to the bridge chip
3) All of that bandwidth is available for IO
4) Nvidia NForce 3-250GB integrates most of the peripherals that you'll
need (Gigabit ethernet, SATA, AGP8X).

There is no PCI Express yet (there will be) but you don't actually need it
because everything that you'll need is already connected to the
hypertransport bus via the highly integrated bridge chips.

So if anything is future proof it's the AMD64.

 

[JK]

The AMD64 runs rings around the P4. I've been running benchmarks on my
Athlon 64 3400+ laptop. In CPU bound applications it's twice as fast as my
2.66GHz Xeon. If there is a lot of IO then it falls to 40% faster but
that's due to the laptop's slow disk (4400RPM vs 7200RPM on the Xeon).

You could get a 7200 rpm hard drive for your notebook. They aren't that
expensive( perhaps around $170 for a 60 gig one).

 

[Noozer]

Thanks. I've decided against PCIx, and know it's just whether I can live
without Hyper-Threading or not. But then, I can always keep my current PC
as well.

You do know that PCIx is *NOT* PCI Extreme (PCIe)

PCIx is a 64bit PCI slot that uses backwards compatible PCI cards.

 

[Synapse Syndrome]

The AMD64 runs rings around the P4. I've been running benchmarks on my
Athlon 64 3400+ laptop. In CPU bound applications it's twice as fast as my
2.66GHz Xeon. If there is a lot of IO then it falls to 40% faster but
that's due to the laptop's slow disk (4400RPM vs 7200RPM on the Xeon). To
put that in MHz terms the 2.2GHz Athlon 64 is equivalent to a 4-5GHz
Xeon depending on the application.

For desktop applications the 939pin A64s have twice the memory bandwidth
of the 754 pin A64 in my laptop which means that they should be even
faster. If you get a motherboard that uses the Nvidia NForce 3-250GB or
Ultra bridge chips you'll get an extremely high performance gigabit
ethernet MAC that's connected directly to the hypertransport bus so it's
not stealing any bandwidth from the PCI slots. The SATA controllers are
also built in so they are stealing any PCI bandwidth either. To top it off
the hypertransport bus that connects the CPU to the bridge chip runs
4GBytes/second and it's solely used for IO. Intel chips access their
memory systems over the same frontside bus as their IO systems, the AMDs
have dedicated on chip memory controllers which are not only higher
performance than Intel's (because the latency is much lower) but also
because they aren't competing for bandwidth with the IO systems.

To sum up,

1) Athlon 64 is much faster than the P4
2) Athlon 64 has more bandwidth to the bridge chip
3) All of that bandwidth is available for IO
4) Nvidia NForce 3-250GB integrates most of the peripherals that you'll
need (Gigabit ethernet, SATA, AGP8X).

There is no PCI Express yet (there will be) but you don't actually need it
because everything that you'll need is already connected to the
hypertransport bus via the highly integrated bridge chips.

So if anything is future proof it's the AMD64.

Thanks. I've decided against PCIx, and know it's just whether I can live
without Hyper-Threading or not. But then, I can always keep my current PC
as well.

ss.

 

[Noozer]

It's quite confusing as all the standards start with "Ex..."

current

BTW, what's PCI Extreme?

PCIe = PCI Extreme = PCI Express = New 1 bit/4 bit/16 bit slots that will be
the new expansion interface card standard.

PCIx = PCI eXtended = Recent update of the PCI bus to allow higher bandwidth
and maintain backwards compatibility. A PCIx card can run up to 133Mhz
(actually current spec is 533mhz) and can use a 64bit bus. Standard PCI
slots are 33Mhz and use a 32bit bus. PCIx cards can usually be used in a
standard PCI slot, at a slower speed.

PCI Specs:
http://www.pcisig.com/specifications/

PCIe:
http://www20.graphics.tomshardware.com/graphic/20040310/pcie-01.html
http://www.microsoft.com/whdc/device/display/PCIe_graphics.mspx#XSLTsection122121120120
http://zone.ni.com/devzone/conceptd.nsf/webmain/70B1D3E338E6F52386256E37006DFDB3
http://www.pcquest.com/content/technology/2004/104040504.asp

PCIx:
http://www.webopedia.com/TERM/P/PCI_X.html
http://www.atmarkit.co.jp/fsys/special/001iasvr_basic/pcix_slots.jpg
http://www.pmasia.com/image/pcix64.jpg

Google is your friend!

 

[Synapse Syndrome]

You do know that PCIx is *NOT* PCI Extreme (PCIe)

PCIx is a 64bit PCI slot that uses backwards compatible PCI cards.

I wasn't talking about PCI-X.

I have seen PCI-Express refered to as PCIe and PCx.

ss.

 

[Synapse Syndrome]

You do know that PCIx is *NOT* PCI Extreme (PCIe)

PCIx is a 64bit PCI slot that uses backwards compatible PCI cards.

BTW, what's PCI Extreme?

ss.

 

[Synapse Syndrome]

It's quite confusing as all the standards start with "Ex..."




PCIe = PCI Extreme = PCI Express = New 1 bit/4 bit/16 bit slots that will be
the new expansion interface card standard.

PCIx = PCI eXtended = Recent update of the PCI bus to allow higher bandwidth
and maintain backwards compatibility. A PCIx card can run up to 133Mhz
(actually current spec is 533mhz) and can use a 64bit bus. Standard PCI
slots are 33Mhz and use a 32bit bus. PCIx cards can usually be used in a
standard PCI slot, at a slower speed.

PCI Specs:
http://www.pcisig.com/specifications/

PCIe:
http://www20.graphics.tomshardware.com/graphic/20040310/pcie-01.html

#
Yes, thank you very much for the explanation. While it was very interesting
it wasn't anything I didn't already know. Just read the first paragraph of
the link you provided above.

ss.

 

[Noozer]

PCI Specs:

#
Yes, thank you very much for the explanation. While it was very interesting
it wasn't anything I didn't already know. Just read the first paragraph of
the link you provided above.

: )

Unfortunately it was the last that I found. Put it first because it was the
most relevant.

Hopefully a few folks read it and will help with the confusion between PCIx
and PCIe.

 

[General Schvantzkoph]

Thanks. I've decided against PCIx, and know it's just whether I can
live without Hyper-Threading or not. But then, I can always keep my
current PC as well.

ss.

Hyperthreading doesn't work very well in the P4, you aren't giving up
anything. Hyperthreading makes sense for some supercomputer architectures
but it doesn't work well for general purpose CPUs. If you design a
processor specifically for hyperthreading, like Burton Smith's machines at
Tera (Now Cray since they bought Cray Research from SGI), you can build
much deeper and simpler pipelines because you can eliminate the dependency
checks and interlocks between most of the stages. Smiths machines had as
many as 128 threads in each processor. The effective performance of each
virtual processor was feable but there were many thousands of virtual
processors. I don't know how well these machines worked or even if they
were ever shipped. The NSA paid for the basic research and one assumes
that if any were actually built they were the customer (I'm guessing that
they must have bought a number of machines because Tera/Cray is the one
remaining stand alone supercomputer company and as far as I know no one
else has bought a system from them).

In the P4 hyperthreading was Intel's attempt to get back some of the
performance they lost by overpipelining their CPUs. The pipelines in the
P4 were over 20 stages deep, vs 11 in the PIII or the AMD CPUs. The
advantage of very deep pipelines is that it allowed for higher clock
rates. The problem with long pipelines is that if you mispredict a branch
then you end up throwing away a lot of instructions. Normal computer code
has a lot of branches so the P4 spends a lot of it's time doing useless
work. Hyperthreading helps because it effectively shortens the pipeline.
If you have two streams going down a 20 stage pipe then each stream sees a
10 stage virtual pipeline. If a branch is mispredicted then you throw away
10 instructions not 20. That's the upside. The downside is that two
unrelated streams are sharing the same cache, so the cache is effectively
half as big. The same is true for registers. The P4 has a large pool of
registers which it renames on the fly (I think the number is 127). When it
is in hyperthreading mode each thread is only allowed to use half of the
registers. The result is that individual threads run slower in
hyperthreading mode (my measurements are 25% slower when a single thread
is executing on a Xeon that has hyperthreading enabled). The overall
throughput of the system does go up but not by much, my measurements were
10%. The only place where the P4 wins is on heavily vectorized code (i.e.
when the multimedia instructions are used). The high clock rate means that
the multimedia instructions are very fast. The kernels of that kind of
code are design to have very few branches and most of those are loops
which are easy to predict (i.e. for(i=0;i<100;i++) will branch in the same
direction for 100 out of 101 tests). However the rest of the program still
runs slower so the net gain is not as high as you would expect.

The high clock rate strategy worked for Intel for a long time because most
people think that clock rate is the same thing as CPU speed. AMD had to
counter with their unfortunately named PR ratings. The PR ratings are
honest but the name sounds like Public Relations rather than Processor
Rating so it always had a fishy sound to it. Intel ran into a barrier
this year with heat and process. The 90nm P4s, which have even deeper
pipelines, are significantly hotter than the 130nm parts and are also a
little slower. They haven't been able to push the clock rate up very much.
To top it off the Pentium M, which has a shallow pipeline and runs much
cooler, is much faster than the P4 on a clock for clock basis. Intel is
now switching to a version of AMDs PR ratings so you won't be hearing
about clock rates anymore (you'll have to read the fine print).

So bottom line, forget about hyperthreading. What you want is a fastest
CPU which is the Athlon 64 family.

 

[Synapse Syndrome]

Hyperthreading doesn't work very well in the P4, you aren't giving up
anything. Hyperthreading makes sense for some supercomputer architectures
but it doesn't work well for general purpose CPUs. If you design a
processor specifically for hyperthreading, like Burton Smith's machines at
Tera (Now Cray since they bought Cray Research from SGI), you can build
much deeper and simpler pipelines because you can eliminate the dependency
checks and interlocks between most of the stages. Smiths machines had as
many as 128 threads in each processor. The effective performance of each
virtual processor was feable but there were many thousands of virtual
processors. I don't know how well these machines worked or even if they
were ever shipped. The NSA paid for the basic research and one assumes
that if any were actually built they were the customer (I'm guessing that
they must have bought a number of machines because Tera/Cray is the one
remaining stand alone supercomputer company and as far as I know no one
else has bought a system from them).

In the P4 hyperthreading was Intel's attempt to get back some of the
performance they lost by overpipelining their CPUs. The pipelines in the
P4 were over 20 stages deep, vs 11 in the PIII or the AMD CPUs. The
advantage of very deep pipelines is that it allowed for higher clock
rates. The problem with long pipelines is that if you mispredict a branch
then you end up throwing away a lot of instructions. Normal computer code
has a lot of branches so the P4 spends a lot of it's time doing useless
work. Hyperthreading helps because it effectively shortens the pipeline.
If you have two streams going down a 20 stage pipe then each stream sees a
10 stage virtual pipeline. If a branch is mispredicted then you throw away
10 instructions not 20. That's the upside. The downside is that two
unrelated streams are sharing the same cache, so the cache is effectively
half as big. The same is true for registers. The P4 has a large pool of
registers which it renames on the fly (I think the number is 127). When it
is in hyperthreading mode each thread is only allowed to use half of the
registers. The result is that individual threads run slower in
hyperthreading mode (my measurements are 25% slower when a single thread
is executing on a Xeon that has hyperthreading enabled). The overall
throughput of the system does go up but not by much, my measurements were
10%. The only place where the P4 wins is on heavily vectorized code (i.e.
when the multimedia instructions are used). The high clock rate means that
the multimedia instructions are very fast. The kernels of that kind of
code are design to have very few branches and most of those are loops
which are easy to predict (i.e. for(i=0;i<100;i++) will branch in the same
direction for 100 out of 101 tests). However the rest of the program still
runs slower so the net gain is not as high as you would expect.

The high clock rate strategy worked for Intel for a long time because most
people think that clock rate is the same thing as CPU speed. AMD had to
counter with their unfortunately named PR ratings. The PR ratings are
honest but the name sounds like Public Relations rather than Processor
Rating so it always had a fishy sound to it. Intel ran into a barrier
this year with heat and process. The 90nm P4s, which have even deeper
pipelines, are significantly hotter than the 130nm parts and are also a
little slower. They haven't been able to push the clock rate up very much.
To top it off the Pentium M, which has a shallow pipeline and runs much
cooler, is much faster than the P4 on a clock for clock basis. Intel is
now switching to a version of AMDs PR ratings so you won't be hearing
about clock rates anymore (you'll have to read the fine print).

So bottom line, forget about hyperthreading. What you want is a fastest
CPU which is the Athlon 64 family.

OkaY!. You sold me..

ss.