4 Hardware bandwidth questions. (Related to ram,pci-e,cpu,gpu,sli).

[Skybuck Flying]

Hello fellow netcitizens I have some hardware bandwidth questions for you:

Q1: Some websites specifications say:

1. The Athlon X2 4800 processor is limited to 6.4 Gigabyte/sec of memory
bandwidth.
2. Intel Pentium Extreme Edition processor is limited to 6.4 Gigabyte/sec of
memory bandwidth.

Case 1. Does this mean 3.2 Gigabyte/sec in both directions ?

Case 2. Or is it possible to have 6.4 Gigabyte/sec going in one direction ?

Q2: Some website specifications say (more or less):

PCI-E 16x slot is 8 Gigabyte/sec, 4 Gigabyte/sec IN and 4 Gigabyte/Sec out
PCI-E 8x slot is 4 Gigabyte/sec, 2 Gigabyte/sec IN and 2 Gigabyte/Sec out
PCI-E 4x slot is 2 Gigabyte/sec, 1 Gigabyte/sec IN and 1 Gigabyte/Sec out
PCI-E 2x slot is 1 Gigabyte/sec, 512 Megabyte/sec IN and 512 Megabyte/Sec
out
PCI-E 1x slot is 512 Megabyte/sec, 256 Megabyte/sec IN and 256 Megabyte/Sec
out

Case 3. So does this mean that PCI-E 16x slot is limited to 4 Gigabyte/sec
in one direction ?

Case 4. Or is it possible to go 8 Gigabyte/sec in one direction ?

I'll assume Case 2 and Case 3 to be true.

Furthermore it seems current SLI motherboards when in SLI mode,
turn the two PCI-E 16x slots into PCI-E 8x slots.

So the bandwidth that is available in SLI mode is the same as in SINGLE
mode.

Anyway in both cases this would mean a maximum input for the graphics cards
of:

Single: 4 Gigabyte/sec Input.

SLI: 2 Gigabyte/sec + 2 Gigabyte/sec = 4 Gigabyte/sec Input.

So that would leave 2.4 Gigabyte/sec for output.
(from graphics card back to RAM)

Single: 2.4 Gigabyte/sec Output

SLI: 1.2 Gigabyte/sec + 1.2 Gigabyte/sec = 2.4 Gigabyte/sec Output.

So if I were to buy such an SLI motherboard it would be pointless to buy a
graphics card which can handle more than 2 gigabyte/sec of input and more
than 1.2 gigabyte/sec of output.

Q3: How much bandwidth can the new Nvidia GTX 7800 process for input and
generate for output ?

Q4: How much bandwidth can the processors handle ?

Alternatively I could wait for new motherboards to come out with FULL
BANDWIDTH PCI-E 16x slots.

Then the picture would look like this:

Single: 4 Gigabyte/sec Input.

SLI: 4 Gigabyte/sec + 4 Gigabyte/sec = 8 Gigabyte/sec Input.

BUT this is now limited by the the processor/memory controller: 6.4
Gigabyte/sec

SLI: 3.2 Gigabyte/sec + 3.2 Gigabyte/sec = 6.4 Gigabyte/sec Input.

That would leave no room for output... that s bad.

So better could be:

SLI: 2 Gigabyte/sec + 2 Gigabyte/sec = 4 Gigabyte/sec Input.

SLI: 1.2 Gigabyte/sec + 1.2 Gigabyte/sec = 2.4 Gigabyte/sec Output/

So I think I must come to the conclusion that with this processor/memory
controller/bandwidth limitation waiting/buying this new motherboard will
solve absolutely nothing. Since the bottleneck is in the processor/memory
controller. (I think sometimes this is called the northbridge chip ? )

Anyway these dual core processors are very expensive.. so just upgrading
them in the future for a little bit more bandwidth seems crazy.

My conclusion is as follows:

1. If the single card (the gtx 7800) is able to handle 6.4 Gigabyte/sec of
input + output then I don't need SLI

2. Otherwise I could buy into SLI as soon as the price has dropped to a
certain point for the secondary graphics card.

The performance increase for SLI is:
( ( Final Performance - Original Performance ) / Original Performance ) *
100%
( ( 6.4 - 4.0 ) / 4.0 ) * 100% = 60%

Let's see the current price is $500 dollars.

So the price of the future card should only be 60% of the price which it is
now.

$5 * 60% = $300 dollars

Alternative way of calculating:

The current performance is 100% procent.

The future performance will be 160% procent.

So the future total price should not exceed 160 * $5 dollars = $800 total
dollars

We might already have payed $500 dollars so an investment of $300 dollars is
warrented to get the same percentagual increase in performance
hehehehehehhihihihi.

Ofcourse by then much better cards might have come out with even superior
performance... though our current motherboards won't be able to benefit from
them, simply because the bottleneck of 4 Gigabyte/sec
in and 4 Gigabyte/sec out... but that would exceed the main memory/ram of
6.4 Gigabyte... so finally we more or less limited to 3.2 Gigabyte/sec for
current motherboards/processors/chipsets

Bye,
Skybuck.

 

[Phil Weldon]

'Skybuck Flying crossposted some questions to way to many newsgroups:
| Hello fellow netcitizens I have some hardware bandwidth questions for you:
..
..
_____

A memory bus is bidirectional; data can move in two directions, but NOT AT
THE SAME TIME.

Memory bandwidths are given in PEAK capacity. No current memory modules can
come even close to delivering data as fast as current CPUs can use the data.

No peripherials currently available (including display adapters) can use
anything near the PEAK capacity of ANY of the newer buses for which they are
designed. Even ATA133 can't be maxed out.

Do a lot more reading and reconsider the rest of your questions.

Phil Weldon

| Hello fellow netcitizens I have some hardware bandwidth questions for you:
|
| Q1: Some websites specifications say:
|
| 1. The Athlon X2 4800 processor is limited to 6.4 Gigabyte/sec of memory
| bandwidth.
| 2. Intel Pentium Extreme Edition processor is limited to 6.4 Gigabyte/sec
of
| memory bandwidth.
|
| Case 1. Does this mean 3.2 Gigabyte/sec in both directions ?
|
| Case 2. Or is it possible to have 6.4 Gigabyte/sec going in one direction
?
|
| Q2: Some website specifications say (more or less):
|
| PCI-E 16x slot is 8 Gigabyte/sec, 4 Gigabyte/sec IN and 4 Gigabyte/Sec out
| PCI-E 8x slot is 4 Gigabyte/sec, 2 Gigabyte/sec IN and 2 Gigabyte/Sec out
| PCI-E 4x slot is 2 Gigabyte/sec, 1 Gigabyte/sec IN and 1 Gigabyte/Sec out
| PCI-E 2x slot is 1 Gigabyte/sec, 512 Megabyte/sec IN and 512 Megabyte/Sec
| out
| PCI-E 1x slot is 512 Megabyte/sec, 256 Megabyte/sec IN and 256
Megabyte/Sec
| out
|
| Case 3. So does this mean that PCI-E 16x slot is limited to 4 Gigabyte/sec
| in one direction ?
|
| Case 4. Or is it possible to go 8 Gigabyte/sec in one direction ?
|
| I'll assume Case 2 and Case 3 to be true.
|
| Furthermore it seems current SLI motherboards when in SLI mode,
| turn the two PCI-E 16x slots into PCI-E 8x slots.
|
| So the bandwidth that is available in SLI mode is the same as in SINGLE
| mode.
|
| Anyway in both cases this would mean a maximum input for the graphics
cards
| of:
|
| Single: 4 Gigabyte/sec Input.
|
| SLI: 2 Gigabyte/sec + 2 Gigabyte/sec = 4 Gigabyte/sec Input.
|
| So that would leave 2.4 Gigabyte/sec for output.
| (from graphics card back to RAM)
|
| Single: 2.4 Gigabyte/sec Output
|
| SLI: 1.2 Gigabyte/sec + 1.2 Gigabyte/sec = 2.4 Gigabyte/sec Output.
|
| So if I were to buy such an SLI motherboard it would be pointless to buy a
| graphics card which can handle more than 2 gigabyte/sec of input and more
| than 1.2 gigabyte/sec of output.
|
| Q3: How much bandwidth can the new Nvidia GTX 7800 process for input and
| generate for output ?
|
| Q4: How much bandwidth can the processors handle ?
|
| Alternatively I could wait for new motherboards to come out with FULL
| BANDWIDTH PCI-E 16x slots.
|
| Then the picture would look like this:
|
| Single: 4 Gigabyte/sec Input.
|
| SLI: 4 Gigabyte/sec + 4 Gigabyte/sec = 8 Gigabyte/sec Input.
|
| BUT this is now limited by the the processor/memory controller: 6.4
| Gigabyte/sec
|
| SLI: 3.2 Gigabyte/sec + 3.2 Gigabyte/sec = 6.4 Gigabyte/sec Input.
|
| That would leave no room for output... that s bad.
|
| So better could be:
|
| SLI: 2 Gigabyte/sec + 2 Gigabyte/sec = 4 Gigabyte/sec Input.
|
| SLI: 1.2 Gigabyte/sec + 1.2 Gigabyte/sec = 2.4 Gigabyte/sec Output/
|
| So I think I must come to the conclusion that with this processor/memory
| controller/bandwidth limitation waiting/buying this new motherboard will
| solve absolutely nothing. Since the bottleneck is in the processor/memory
| controller. (I think sometimes this is called the northbridge chip ? )
|
| Anyway these dual core processors are very expensive.. so just upgrading
| them in the future for a little bit more bandwidth seems crazy.
|
| My conclusion is as follows:
|
| 1. If the single card (the gtx 7800) is able to handle 6.4 Gigabyte/sec of
| input + output then I don't need SLI
|
| 2. Otherwise I could buy into SLI as soon as the price has dropped to a
| certain point for the secondary graphics card.
|
| The performance increase for SLI is:
| ( ( Final Performance - Original Performance ) / Original Performance ) *
| 100%
| ( ( 6.4 - 4.0 ) / 4.0 ) * 100% = 60%
|
| Let's see the current price is $500 dollars.
|
| So the price of the future card should only be 60% of the price which it
is
| now.
|
| $5 * 60% = $300 dollars
|
| Alternative way of calculating:
|
| The current performance is 100% procent.
|
| The future performance will be 160% procent.
|
| So the future total price should not exceed 160 * $5 dollars = $800 total
| dollars
|
| We might already have payed $500 dollars so an investment of $300 dollars
is
| warrented to get the same percentagual increase in performance
| hehehehehehhihihihi.
|
| Ofcourse by then much better cards might have come out with even superior
| performance... though our current motherboards won't be able to benefit
from
| them, simply because the bottleneck of 4 Gigabyte/sec
| in and 4 Gigabyte/sec out... but that would exceed the main memory/ram of
| 6.4 Gigabyte... so finally we more or less limited to 3.2 Gigabyte/sec for
| current motherboards/processors/chipsets
|
| Bye,
| Skybuck.
|
|

 

[Skybuck Flying]

I have already googled extensively for the words bandwidth, cpu, gpu, ram,
diagrams, netburst, hyper transport, maximum, graphics etc.

Not once have I encountered a website which states the ammount of bandwidth
that a cpu, gpu, ram chip or transport technology can practically process or
generate.

The only things I have encountered are:

1. Occording to you apperently theoretical maximum bandwidths.

2. PC/System benchmarks, which ofcourse depend on many factors and
combinations and are therefore bogus.

You claim to have knowledge of practical limitations of these devices.

I claim that you are full of bullshit since this information is not
available on the internet, so there is after all nothing to read.

Ofcourse you are free to prove me wrong =D

Bye,
Skybuck.

 

[Phil Weldon]

'Skybuck Flying' wrote, in part:
|I have already googled extensively for the words bandwidth, cpu, gpu, ram,
| diagrams, netburst, hyper transport, maximum, graphics etc.
|
| Not once have I encountered a website which states the ammount of
bandwidth
| that a cpu, gpu, ram chip or transport technology can practically process
or
| generate.
_____

Google + websites do not necessarily = learning.
Websites are not the only place one can read.
While googling and surfing you might also read up on
how to post to a newsgroup
why it is a good idea to read at least some of the posts in newsgroups
before posting
why it is a good idea to trim quotes when posting
why it is a good idea to limit crossposting.

And you might also consider not asking a question if you don't want an
answer. A different attitude might invite more answers to your questions.

Phil Weldon

|I have already googled extensively for the words bandwidth, cpu, gpu, ram,
| diagrams, netburst, hyper transport, maximum, graphics etc.
|
| Not once have I encountered a website which states the ammount of
bandwidth
| that a cpu, gpu, ram chip or transport technology can practically process
or
| generate.
|
| The only things I have encountered are:
|
| 1. Occording to you apperently theoretical maximum bandwidths.
|
| 2. PC/System benchmarks, which ofcourse depend on many factors and
| combinations and are therefore bogus.
|
| You claim to have knowledge of practical limitations of these devices.
|
| I claim that you are full of bullshit since this information is not
| available on the internet, so there is after all nothing to read.
|
| Ofcourse you are free to prove me wrong =D
|
| Bye,
| Skybuck.
|
|| >
| >
|
|

 

[Skybuck Flying]

Here is a nice start:

Technical specifications of GeForce 7800:

http://www.nvidia.com/page/specs_gf7800.html

Graphics Bus Technology: PCI Express
Memory Interface: 256-bit
Memory Bandwidth (GB/sec.): 38.4
Fill Rate (billion pixels/sec.): 10.32
Vertices/sec. (million): 860
Pixels per clock (peak): 24
RAMDACs (MHz): 400

In particular notice this line:

Memory Bandwidth (GB/sec.): 38.4

Is this a theoretical or practically limitation ?

I would guess a practical since the rest looks pretty practical too.

However this information could be misleading.

What is ment with memory bandwidth in this case ?

Case 1: Main/PC/System RAM <=> System Bus <=> Graphics Card

or

Case 2: Onboard graphics card memory/ram communication.

My guess would be 2 !!!!! ( Because PCI express can't even handle that much
bandwidth at the present time)

Which is still pretty ****ing amazing

Here we are with an Athlon X2 4800 dual core which is able to "only" manage
to pump around 6.4 GB/sec...

Gjee...

Maybe I am wrong and maybe it's case 2 =D <- that would be cool =D

Bye,
Skybuck.

 

[Skybuck Flying]

You might be interested in getting a clue to what you're talking about

Bye, Bye,
Mister Trolly

 

[Skybuck Flying]

Once again I am happy to disappoint you:

It's rare, here is ****ing bandwidth test, instead of other real world not
so real fake bullshit tests:

http://www.legionhardware.com/document.php?id=455&p=3

Jiiiiiiiiiiiiiiiiiippppppppppppeeeeeeeeeeeeekwaaaaaaaajeeeeeeeeee
motherfucker

Look at that bandwidth fly by !!!!
wwwwwwwoooooooooooeeeeeeeeeehhhhhhhhhoooooooeeeeeeewwwwwwwiiiiieeeeeeeeeeeee
eeeeeeeeeeeeeeeeeeeeeee

Well into the multiple gigabytes/sec ! =D

Bye,
Skybuck.

 

[Mac]

'Skybuck Flying crossposted some questions to way to many newsgroups:
| Hello fellow netcitizens I have some hardware bandwidth questions for you:
.
.
_____

A memory bus is bidirectional; data can move in two directions, but NOT AT
THE SAME TIME.

[snip]

Just to clarify, in the Intel architecture, the processor is not connected
to the memory. The processor bandwidth given would probably be the host
bus bandwidth. This is the bus connecting the processor to the so-called
north bridge, which is where the memory controller and PCI bus are.

--Mac

 

[Richard Hopkins]

You might be interested in getting a clue to what you're talking
about

You would do well to go away, reconsider your attitude, taking on the fair
advice that Phil has given you, and then try again, preferably after
apologising for coming across as an ignorant arsehole. Does the saying "Do
not bite the hand that feeds" mean anything to you?
--


Richard Hopkins
Cardiff, Wales, United Kingdom
(replace nospam with pipex in reply address)

The UK's leading technology reseller www.dabs.com

 

[Benjamin Gawert]

Skybuck Flying wrote:

[removed the groups where this is OT]

Q1: Some websites specifications say:

1. The Athlon X2 4800 processor is limited to 6.4 Gigabyte/sec of memory
bandwidth.
2. Intel Pentium Extreme Edition processor is limited to 6.4 Gigabyte/sec of
memory bandwidth.

Case 1. Does this mean 3.2 Gigabyte/sec in both directions ?

No. The FSB of the Pentium4 is bidirectional, being able to do 6.4GB/s
in both directions. The AMD x64 CPUs (Athlon64, Opteron) use
Hypertransport links that are unidirectional...

But then, with intel all processors share the same FSB which is a really
big limitation when having more than one CPU...

Case 2. Or is it possible to have 6.4 Gigabyte/sec going in one direction ?
Right.

Q2: Some website specifications say (more or less):

PCI-E 16x slot is 8 Gigabyte/sec, 4 Gigabyte/sec IN and 4 Gigabyte/Sec out
PCI-E 8x slot is 4 Gigabyte/sec, 2 Gigabyte/sec IN and 2 Gigabyte/Sec out
PCI-E 4x slot is 2 Gigabyte/sec, 1 Gigabyte/sec IN and 1 Gigabyte/Sec out
PCI-E 2x slot is 1 Gigabyte/sec, 512 Megabyte/sec IN and 512 Megabyte/Sec
out
PCI-E 1x slot is 512 Megabyte/sec, 256 Megabyte/sec IN and 256 Megabyte/Sec
out

One PCIe link (PCIe 1x) provides 250MB/s in and out. PCIe 16x can do
4GB/s in and out.

Case 3. So does this mean that PCI-E 16x slot is limited to 4 Gigabyte/sec
in one direction ?

No. 4GB/s in every direction. PCIe is unidirectional...

Case 4. Or is it possible to go 8 Gigabyte/sec in one direction ?
No.

Furthermore it seems current SLI motherboards when in SLI mode,
turn the two PCI-E 16x slots into PCI-E 8x slots.
Right.

So the bandwidth that is available in SLI mode is the same as in SINGLE
mode.

Regarding the overall bandwith, yes.

Anyway in both cases this would mean a maximum input for the graphics cards
of:

Single: 4 Gigabyte/sec Input.

.... and output.

SLI: 2 Gigabyte/sec + 2 Gigabyte/sec = 4 Gigabyte/sec Input.

No. 2GB/s in and out. You can't just add these numbers...

So that would leave 2.4 Gigabyte/sec for output.
(from graphics card back to RAM)
Single: 2.4 Gigabyte/sec Output

SLI: 1.2 Gigabyte/sec + 1.2 Gigabyte/sec = 2.4 Gigabyte/sec Output.

No. 4GB/s for single gfx card and 2GB/s for SLI...

So if I were to buy such an SLI motherboard it would be pointless to buy a
graphics card which can handle more than 2 gigabyte/sec of input and more
than 1.2 gigabyte/sec of output.

No. Even 2GB/s are much more than what even the latest and greatest
games need....

Q3: How much bandwidth can the new Nvidia GTX 7800 process for input and
generate for output ?

The 7800GTX has a native PCIe 16x interface which can handle up to 4GB/s
in and out (16x 250MB/s)...

The internal bandwith between GPU and gfx card RAM is much higher,
though. The 7800GTX has an internal bandwith of around 38GB/s (yes,
gigabytes!) between the GPU and the gfx card memory...

Q4: How much bandwidth can the processors handle ?

see above.

Alternatively I could wait for new motherboards to come out with FULL
BANDWIDTH PCI-E 16x slots.

Won't take long...

Then the picture would look like this:

Single: 4 Gigabyte/sec Input.

.... and output.

SLI: 4 Gigabyte/sec + 4 Gigabyte/sec = 8 Gigabyte/sec Input.

.... and output.

BUT this is now limited by the the processor/memory controller: 6.4
Gigabyte/sec

SLI: 3.2 Gigabyte/sec + 3.2 Gigabyte/sec = 6.4 Gigabyte/sec Input.

That would leave no room for output... that s bad.

Nope. As I said even PCIe8x is much more than current and even future
games need....

So I think I must come to the conclusion that with this processor/memory
controller/bandwidth limitation waiting/buying this new motherboard will
solve absolutely nothing. Since the bottleneck is in the processor/memory
controller.

Nope. The

Anyway these dual core processors are very expensive.. so just upgrading
them in the future for a little bit more bandwidth seems crazy.

My conclusion is as follows:

1. If the single card (the gtx 7800) is able to handle 6.4 Gigabyte/sec of
input + output then I don't need SLI

Apples and oranges. SLI is not there to improve bus performance. SLI is
there to improve rendering speed. Again, the busses are more than fast
enough. The limitation still is the GPU, that's why there are techniques
like SLI...

Benjamin

 

[Skybuck Flying]

You would do well to go away, reconsider your attitude, taking on the fair
advice that Phil has given you, and then try again, preferably after
apologising for coming across as an ignorant arsehole. Does the saying "Do
not bite the hand that feeds" mean anything to you?

I was gonna say: "Go suck his dick"

But it's clear you already doing that lol.

Bye,
Skybuck.

 

[Skybuck Flying]

Hello Benjamin thank you for some answers.

However this thread is ofcourse about future limitations.

What games do now is totally uninteresting.

What games will do in the future say 1,2,3,4,5 even 10 years from now is
much more interesting.

Currently I am not aware of any software or hardware which allows measuring
how much bandwidth travels through the computer at any giving time, playing
games, using applications, etc, whatever.

These tools possibly don't exist that's why people have to fall back on
games, which ofcourse don't push the system to it's limits. Doom 3 is an
example which simple limits the game to 60 fps per second.

Other games might use less textures than the system can handle etc.

Therefore I am more interested in software, maybe like sandra synthetic
benchmarks which push every little piece to it's maximum performance. Still
even these benchmarks might not be suited to test individual components
since the individual component being tested might not reach it's full
potential because of other system limitations.

Knowing each component's true limitation/maximum performance is important
for selecting components to build a PC with.

The longer the PC lasts and keeps up the more valuable it is. A pc which
cant keep up because some individually component is the bottleneck is less
valuable. The problem is how to find out which component is the bottleneck


When simply measuring the following dilemma exists:

1. Is the component the bottleneck ?

or

2. Is the system bottlenecking the component ?



So having some specifications about maximum throughput of each technology
can help.

Also tools to measure the bandwidth to get more insight into the system
performance would help greatly.

Combining these two pieces of information would allow one to pin point/guess
which component might be the limitation factor/bottleneck. Etc.

Apples and oranges. SLI is not there to improve bus performance. SLI is
there to improve rendering speed. Again, the busses are more than fast
enough. The limitation still is the GPU, that's why there are techniques
like SLI...

How do you know the GPU is the limitation ? Do you mean executing
instructions ?

Prove it

Bye,
Skybuck.

 

[John Fields]

I was gonna say: "Go suck his dick"

But it's clear you already doing that lol.

 

[Benjamin Gawert]

Hello Benjamin thank you for some answers.

However this thread is ofcourse about future limitations.

What games do now is totally uninteresting.

What games will do in the future say 1,2,3,4,5 even 10 years from now is
much more interesting.

Ok, then think about on what computers were highend 10 years ago:
Pentium 200MMX with 16MB RAM and one of the first 3D cards like a Matrox
Millenium PCI. And now think about who cares about these computers
today. Exactly. no-one. Almost everything has changed during the last 10
years: busses, processors, memory, gfx cards, and much more...

In 10 years from now certainly no-one gives a **** on PCIe 16x any more...

Currently I am not aware of any software or hardware which allows measuring
how much bandwidth travels through the computer at any giving time, playing
games, using applications, etc, whatever.

These tools possibly don't exist that's why people have to fall back on
games, which ofcourse don't push the system to it's limits. Doom 3 is an
example which simple limits the game to 60 fps per second.

Other games might use less textures than the system can handle etc.

Therefore I am more interested in software, maybe like sandra synthetic
benchmarks which push every little piece to it's maximum performance. Still
even these benchmarks might not be suited to test individual components
since the individual component being tested might not reach it's full
potential because of other system limitations.

What do you want your PC to be: a tool or a penis enlargement? PCIe is
the last thing that is a performance bottleneck. Number one are disk
drives which are slow like hell compared to all other data transfer
processes in a PC. Second is RAM which is still plain slow compared to
what CPUs are able to transfer...

Knowing each component's true limitation/maximum performance is important
for selecting components to build a PC with.

Nope. Especially since just counting numbers like you do is quite silly.
To understand the limitations of a system you need much deeper knowledge
than just a few figures you read from a website. And it heavily depends
on the applications that are used, and the data that is processed, and a
lot of other variables that simply make it impossible to give something
like a universal rule...

And it's also important to understand that benchmarks like the ones
you're looking for are just that: benchmarks. And the only thing they
can tell you is how a computer performs at that benchmark. They _don't_
tell you how this PC will perform with real applications...

The longer the PC lasts and keeps up the more valuable it is. A pc which
cant keep up because some individually component is the bottleneck is less
valuable. The problem is how to find out which component is the bottleneck

No matter what you buy today, it _will_ get outdated. And no matter what
you buy, even the ultra-best and finest components that you can buy for
money today will perform like shit compared to what will be available in
any el-cheapo PC in 3, 4, or 5 years from now...

How do you know the GPU is the limitation ?

Because SLI is made to increase the rendering performance, not because
the bus limits throughput...

Do you mean executing
instructions ?

Prove it

I don't have to. Anyone who is at least at medium knowledge in computer
hardware should know that. You asked a question and got an response.
What you make from it is your part. But there should be enough sources
around the web which will explain you where the bottlenecks in todays
computers are and where not...

Benjamin

 

[Skybuck Flying]

Ok, then think about on what computers were highend 10 years ago:

Ok, windows 95 era.

Pentium 200MMX with 16MB RAM and one of the first 3D cards like a Matrox
Millenium PCI.

Great computer. Internet, word processing, even some shooters which haven't
changed that much... aka doom3 aka quake 4

And now think about who cares about these computers
today.

People who want to internet and word process ?

Exactly. no-one.

Lol, are you the all knowing god ?

I think a poor person in some under developed country would be happy with
such a computer.

(Or people fed up with cleaning out dust from the dust sucking pc's nowadays
LOL)

Almost everything has changed during the last 10
years: busses, processors, memory, gfx cards, and much more...

Gje, when will those bastards get it right huh ?

In 10 years from now certainly no-one gives a **** on PCIe 16x any more...

Most people don't give a **** today since their clueless ?

What do you want your PC to be: a tool or a penis enlargement? PCIe is
the last thing that is a performance bottleneck.

I hope so, that would be great.

Number one are disk drives which are slow like hell compared to all other data transfer
processes in a PC.

True, access time hasn't changed at all the last 10 years.

Throughput however has become quite fast...

Second is RAM which is still plain slow compared to
what CPUs are able to transfer...

Hmm.. I like to think of RAM being fast.

How fast do you think CPU can generate data ?

Nope. Especially since just counting numbers like you do is quite silly.

I think it's quite smart to start at the theoretical limitations and proceed
downwards from there.

To understand the limitations of a system you need much deeper knowledge
than just a few figures you read from a website.

I read a lot more than just some numbers

And it heavily depends on the applications that are used,

Then stop using those applications to measure your system's performance.

Use a synthetic benchmark to test the true performance of your hardware.

and the data that is processed, and a
lot of other variables that simply make it impossible to give something
like a universal rule...

You should seperate hardware performance from software performance since
those two things are two completely seperate things.

And it's also important to understand that benchmarks like the ones
you're looking for are just that: benchmarks.

No there are different benchmarks.

I in particular like the benchmarks which drive every component to it's
maximum.

And the only thing they can tell you is how a computer performs at that

benchmark.

The idea is ofcourse to use a good benchmark which will show the maximum
performance of the hardware.

They _don't_ tell you how this PC will perform with real applications...

Again you are mixing software performance with hardware performance.

If the hardware performance is good and the software performance is bad then
what does that tell you ?

No matter what you buy today, it _will_ get outdated. And no matter what
you buy, even the ultra-best and finest components that you can buy for
money today will perform like shit compared to what will be available in
any el-cheapo PC in 3, 4, or 5 years from now...

The point of this thread is to focus on bottlenecks and raise those necks so
that the computer will keeping performing incredibly well for the coming
1,2,3,4,5 years at least.

Component selection/understanding is what it's all about. Do your job well
today and you'll be laughing at people 5 years down the road because they
bought a PC off the shelf with a 5000+ CPU but have some serious bottleneck
elsewhere without them even knowing about it

Maybe you are wrong.

It seems that future motherboards can simply add PCI-E lanes and increase
bus performance.

SLI simply uses multiple PCI slots and could therefore increase the ammount
of bandwidth flowing to the graphics cards.

SLI is there to improve rendering speed.

Lol, you funny, what do you think is needed for rendering ? Exactly data !
That data has to be transferred !
Thus the bandwidth bottleneck is born

Again, the busses are more than fast enough.

Oops, now you are just dead wrong.

Just install poorly written game x out of a million and watch your pc crawl
to a halt

In other words, 25 textures of 1024x1024x32 bits colors being pushed through
the system at 70 frames per second. That's 25 * 70 * 4 MB = 100 * 70 = 7
GB/sec.

And we haven't talked about all other things that need to be transferred
etc...

Even the 7 GB/sec the athlon can't do.

The make matters worse a single card can't pull that kind of data simply
because PCI-E 16x slot is limited to 4 GB/sec.

So these figures are worrieing.

So maybe the following statement could be true:

If I choose a pentium chip over an athlon chip then the pentium chip might
have lower frames rates in present games... but the pentium chip might
achieve the same frame rate in future games simply because it can sustain a
higher bandwidth.

The athlon chip simply can't provide enough bandwidth and therefore an
athlon based system will simply hit the bottleneck much sooner and cause
games to crawl.

So as you can see it's very interesting to get some facts/measurements about
all this and not just somebody on a newsgroup claiming that bussess are fast
enough

I was hoping to use nvidia's NVPerfMon to do some measurements on some of my
favorite opengl games on my current computer and to figure out what the
bottlenecks are for these games, which could be extrapolated onto future
system requirements

The limitation still is the GPU, that's why there are techniques

Because SLI is made to increase the rendering performance, not because
the bus limits throughput...

Why use SLI in the first place ?

Why not simply put two GPU's on one card

Maybe Scalable Link Interface is like two network cards: it increases the
bandwidth.

Some motherboard manufacturers are coming out with two "full bandwidth"
pci-e 16x slots.

Ofcourse this could only be marketing bullshit or maybe there is some thruth
in it.

The current bottleneck is not PCI-E or SLI but the bottleneck is the
CPU<->RAM connection which can only deliver 6.4 GB/sec. (As I tried to
explain in one of my previous post).

The question is thus if Athlons/SLI/Nforce4/Geforces are worth their money
price/performance wise.

Euhm what are our choices:

1. Buy into Athlons/SLI/Nforce4/2 Geforces and then be bandwidth limited.

or

2. Buy into Pentium/????

At this point I am not sure if the bandwidth limitation is Nforce4 chipset
specific or if it's only athlon cpu specific. (Think of integrated memory
controller etc, northbridge chip etc)

The northbridge chip is normally part of the chipset I believe so it could
be Nforce4 chipset specific... that could mean a new/different chipset is
needed which doesn't have this main memory bandwidth limitation to fully
justify SLI/double graphics cards etc...

It's quite interesting to see if their is or is not an alternative competing
technology from intel or ati

However ati has buggy drivers so that rules out ati, which means intel
remains... though intel is said to have power hungry and overheating cpu's
so the athlon/sli/nvidia deal seems best at the moment =D

I don't have to. Anyone who is at least at medium knowledge in computer
hardware should know that. You asked a question and got an response.
What you make from it is your part. But there should be enough sources
around the web which will explain you where the bottlenecks in todays
computers are and where not...

Cheap.

You should really try to figure stuff out yourself and if you have the
chance test it... otherwise you run the risk of marketing people screwing
you over lol =D

Bye,
Skybuck

 

[Benjamin Gawert]

Skybuck Flying wrote:

[due to the amount of nonsense I answered only the most anoying parts]

Ok, windows 95 era.




Great computer. Internet, word processing, even some shooters which haven't
changed that much... aka doom3 aka quake 4

Maybe these shooters didn't change much regarding game content. But they
changed a lot when it comes to system ressources...

People who want to internet and word process ?

Yeah, right. You definitely want to do internet with an outdated OS that
doesn't get updated any more like Windows95, because only there you can
be sure to catch _every_ bit of malware that's around...

Lol, are you the all knowing god ?

I think a poor person in some under developed country would be happy with
such a computer.

Sure. Tell the people that are dying because of hunger or diseases
because they are so poor that they can't afford food and medicine that
everything is good as long as they have that old Pentium200 running
Windows95. It's not that they don't have any bigger problems...

Besides that, even in development countries, such a system will be quite
useless because even there it catches all worms that are floating
around. But yeah, we definitely need more unsecure computers that are
used as SPAM sources...

Gje, when will those bastards get it right huh ?




Most people don't give a **** today since their clueless ?

Or perhaps they just know it better than you?

data transfer



True, access time hasn't changed at all the last 10 years.

Throughput however has become quite fast...

Yeah, right. ~60MB/s is sooo fast. And you're whining about PCIe being a
bottleneck...

Hmm.. I like to think of RAM being fast.

Wrong thought.

How fast do you think CPU can generate data ?

The old Pentium4 1.5GHz that sits here besides my desk does around
12GB/s between CPU and L1 cache and ~10GB/s between CPU and L2 cache...

BTW. that's the reason why processors do have cache: simply because RAM
is so slow...

I think it's quite smart to start at the theoretical limitations and proceed
downwards from there.

Nope, it's useless, especially as a end user like you that only has the
choice of a certain amount of components that are available on the
market. As long as you don't do electronics engineering such discussions
are worthless...

I read a lot more than just some numbers

Yes, you read a lot, but you obviously are lacking the technical
background to fully understand what's really going on inside the
computer. Sadly, most websites won't help you there...

Then stop using those applications to measure your system's performance.

Use a synthetic benchmark to test the true performance of your hardware.

Synthetic benchmarks are called "synthetic" because they have no
relation to real-world applications. But the latter are the things that
are important, because most people buy a computer to work with real
applications and not for running synthetic benchmarks...

You should seperate hardware performance from software performance since
those two things are two completely seperate things.

Nope, they are not. Software performance depends on the hardware and
vice versa...

No there are different benchmarks.

I in particular like the benchmarks which drive every component to it's
maximum.

Fine. But that's really only useful if you see your PC as a penis
protesis...

Component selection/understanding is what it's all about. Do your job well
today and you'll be laughing at people 5 years down the road because they
bought a PC off the shelf with a 5000+ CPU but have some serious bottleneck
elsewhere without them even knowing about it

So, then tell us what PC one should have choosen 5 years ago to be able
to lough about people that have a current system?

Maybe you are wrong.

No. Maybe you should start reading Nvidias SLI documentation...

It seems that future motherboards can simply add PCI-E lanes and increase
bus performance.

Not ad infinitum.

SLI simply uses multiple PCI slots and could therefore increase the ammount
of bandwidth flowing to the graphics cards.

No, since bandwith isn't the problem...

Lol, you funny, what do you think is needed for rendering ? Exactly data !
That data has to be transferred !
Thus the bandwidth bottleneck is born

BS. You really should learn some basics first. Of course generating 3D
visuals needs huge amounts of data. But these data are not generated by
the CPU but by the GPU. The amount of data that is floating from the CPU
is much smaller than what leaves the GPU. That's why the GPUs have such
fast connections to their on board memory, and why even AGP4x is more
than fast enough...

Oops, now you are just dead wrong.

Just install poorly written game x out of a million and watch your pc crawl
to a halt

In other words, 25 textures of 1024x1024x32 bits colors being pushed through
the system at 70 frames per second. That's 25 * 70 * 4 MB = 100 * 70 = 7
GB/sec.

Again you prove your lack of knowledge. Textures are only loaded once in
the gfx card RAM, not again and again. And of course you're totally
ignoring the fact that the _GPU_ does all the calculations, not the CPU.
The CPU doesn't have to send the content pixel for pixel to the gfx
processor. The CPU send raw position and processing data, and the GPU
does the processing....

If I choose a pentium chip over an athlon chip then the pentium chip might
have lower frames rates in present games... but the pentium chip might
achieve the same frame rate in future games simply because it can sustain a
higher bandwidth.

The athlon chip simply can't provide enough bandwidth and therefore an
athlon based system will simply hit the bottleneck much sooner and cause
games to crawl.

So as you can see it's very interesting to get some facts/measurements about
all this and not just somebody on a newsgroup claiming that bussess are fast
enough

Well, I _know_ that the busses are fast enough. Which btw every somewhat
reliable source confirms...

Why use SLI in the first place ?

SLI is technology from 1997, made by 3DFx and bought by Nvidia. The
reason SLI is quite rare is that most games simply don't scale linear
with the number of gfx cores. So having two GPUs doesn't mean you get 2x
the performance of a single card...

Why not simply put two GPU's on one card

Asus does that...

Maybe Scalable Link Interface is like two network cards: it increases the
bandwidth.

Again, read the documentation first...

Cheap.

You should really try to figure stuff out yourself and if you have the
chance test it... otherwise you run the risk of marketing people screwing
you over lol =D

Well, I'm already very deep into that stuff in my job so I don't need
any marketing droids or geeks to get my informations from...

EOD

Benjamin

 

[Ed Forsythe]

I agree Phil but it's obvious that this cretin is a troll so why feed him?

 

[Skybuck Flying]

Skybuck Flying wrote:

[due to the amount of nonsense I answered only the most anoying parts]

Good, that means you getting tired of your own non-sense

Maybe these shooters didn't change much regarding game content. But they
changed a lot when it comes to system ressources...

So what the classics are better anyway LOL.

Yeah, right. You definitely want to do internet with an outdated OS that
doesn't get updated any more like Windows95, because only there you can
be sure to catch _every_ bit of malware that's around...

Simply patch it, update it and it won't be that bad.

Sure. Tell the people that are dying because of hunger or diseases
because they are so poor that they can't afford food and medicine that
everything is good as long as they have that old Pentium200 running
Windows95. It's not that they don't have any bigger problems...

Besides that, even in development countries, such a system will be quite
useless because even there it catches all worms that are floating
around. But yeah, we definitely need more unsecure computers that are
used as SPAM sources...

Bull, see above.

There are even companies specially in re-using old computer equipment for th
ird world countries.

Or perhaps they just know it better than you?

Nope, I doubt it

Yeah, right. ~60MB/s is sooo fast. And you're whining about PCIe being a
bottleneck...

Dude, where do you get these silly numbers ?

1. First of all my own old PIII 450 mhz is able to read with 180 MB/sec from
the harddisk and that is only the harddisk, go figure !

2. Second of all the sandra benchmark shows a bandwidth of 7 GB/sec. So that
means either sandra is full of bullshit or you are full of bullshit LOL.

Gjee I wonder which case it might be ? Euhhmmm yeah me goes for second
option: you full of bullshit LOL.

Wrong thought.

Right thought, get your facts straight.

The old Pentium4 1.5GHz that sits here besides my desk does around
12GB/s between CPU and L1 cache and ~10GB/s between CPU and L2 cache...

I wouldn't know about that... but you have inspired me to go test everything
to understand my current system's limitations... I am especially interested
in finding out if the system buys is the most limitating factor.

Seeing you being off on all other figures no doubt are you off on these
figures as well

BTW. that's the reason why processors do have cache: simply because RAM
is so slow...

Sure if you say so, NOT

Nope, it's useless, especially as a end user like you that only has the
choice of a certain amount of components that are available on the
market. As long as you don't do electronics engineering such discussions
are worthless...

At least I am not an idiot like you, and yes I can get any component I want
thanks to the internet.

Welcome to the real world, from crawling forth from under your stinking ROCK
lol.

Yes, you read a lot, but you obviously are lacking the technical
background to fully understand what's really going on inside the
computer. Sadly, most websites won't help you there...

It's obvious that you are a clueless troll and I wish you lot's of lack with
it.

Synthetic benchmarks are called "synthetic" because they have no
relation to real-world applications. But the latter are the things that
are important, because most people buy a computer to work with real
applications and not for running synthetic benchmarks...

Sure, synthetic benchmarks are useless.. (NOT ) the only push the system
to it's maximum performance.... gjeee

Nope, they are not. Software performance depends on the hardware and
vice versa...

Sure thing trollyboy, if you say so, NOT

Fine. But that's really only useful if you see your PC as a penis
protesis...

Lol, all trolls think like that.

So, then tell us what PC one should have choosen 5 years ago to be able
to lough about people that have a current system?

HAHAHAHAHAHAHAHAHAHAHAHAHAHA.

Idiots like you ofcourse with systems which only achieve 66 MB/sec
hahahahahahahahaahahahahaha.

No. Maybe you should start reading Nvidias SLI documentation...

Maybe you should give it up troll

Not ad infinitum.

Gje, you think ?

No, since bandwith isn't the problem...

Sure if you say so troll, NOT.

BS. You really should learn some basics first. Of course generating 3D
visuals needs huge amounts of data. But these data are not generated by
the CPU but by the GPU. The amount of data that is floating from the CPU
is much smaller than what leaves the GPU. That's why the GPUs have such
fast connections to their on board memory, and why even AGP4x is more
than fast enough...

Sure, if you say so TROLL, get some facts man, at least be a plausible
troll, LOL.

Again you prove your lack of knowledge. Textures are only loaded once in
the gfx card RAM, not again and again. And of course you're totally
ignoring the fact that the _GPU_ does all the calculations, not the CPU.
The CPU doesn't have to send the content pixel for pixel to the gfx
processor. The CPU send raw position and processing data, and the GPU
does the processing....

Dude, get real, try a fast moving shooters, enough said.

Well, I _know_ that the busses are fast enough. Which btw every somewhat
reliable source confirms...

You know jack squate, good luck with your next bandwidth limited PC

HAHAHAHAHAHAHAHAHAHAHA

What will be next, 100 MB/sec ? LOL !

SLI is technology from 1997, made by 3DFx and bought by Nvidia. The
reason SLI is quite rare is that most games simply don't scale linear
with the number of gfx cores. So having two GPUs doesn't mean you get 2x
the performance of a single card...

Oh boy, did you bother read an actually benchmark this time ? Good for you
LOL.

You still have no clue to how fast it really is.

Asus does that...

Yeah so ?

Again, read the documentation first...

Try getting a clue yourself for a change

Well, I'm already very deep into that stuff in my job so I don't need
any marketing droids or geeks to get my informations from...

LOL,

You just another incompetent clueless idiot claiming to have a job
hahahahahahaha.

I am pretty mcuh done with you, have a nice day

Bye,
Skybuck =D Wieeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee.

 

[Mushr00mhead]

So what is the point of a newsgroup if all you are going to do is bash
on one another. What would be nice is if you guys could support your
arguments with some credible facts. Don't just regurgitate what you
think is correct, give some supporting facts to go with it. As for the
name calling, is it necessary?

 

[Mushr00mhead]

Oh yeah, one more thing. Isn't this newsgroup about game programming
and algorithms? If not, then I guess I need to go find that newsgroup.