Bottlenecks

[Jure Sah]

(e-mail address removed) pravi:

Here is another suggestion: If you have that kind of money, then why
be ordinary.

1) For a few billions $$$; Intel/Amd will create you a nice custom CPU
- complete with self-cooling feature that included 10, 20,... cores
(cost you more if you want more cores on the die).

Multicore system? Windows currently has poor support for even 2 cores,
what about more.

And I doubt that with those few billions of $$$, you'd be able to pull
out of the sky better technologies than are currently developed by the
companies.

I took lessons in digital circuits and believe me, getting together a
CPU with 64 bits of functions (18446744073709551616) and put it all in
one optimized chip is not easy.

2) For a few billions $$$; I'm sure Asus/Abit or whoever would create a
nice MB that would work with the above CPU plus all PCI-Exp/PCI/MEM
slot and all the bell and stuff you ever want.

You've missed PCI64, it has a higher persistent transfer rate than
PCI-Express. But why limit oneself to such standards, if you created
your own hardware, you could plug it directly into the motherboard with
no slot interface and pick a communication frequency yourself.

And yet, there is always the question, if it's so easy, why don't they
do it?

3) For a couple more billion $$$; ATI / Nvidia would put together a 10,
20, 30 top of the line GPU SLI together (instead of 2 SLI now a day) -
make sure the above MB has enough PCI Exp slots for let say 20 graphics
card SLI.

Oh please, don't talk like you were born yesterday. An ISA Creative
sound blaster can be connected to as many other copies of the card as
you can fit to a bus.

The problem of course, is having a CPU with as much communication
bandwidth (for all those graphic cards). There are some physical limits
to how fast you can relay a signal without degrading it and now amount
of money can prevent them.

4) For a few millions $$$; someone will create you a nice case to fit
the MB all the above stuff; along with a nice high power PS (1,000,000
Watt)

Perhaps we would best dedicate a nuclear power plant to the thing.
However, IMHO, a computer that uses as much electricity as a small city
is not much of an ultimate computer at all.

5) For a billion or two; someone will make you a couple Tregabite of
RAM (make sure your MB support unlimited memory).

No problem. However all memory needs to be:
1. Fast enough... if you have a lot of memory, it takes a lot of CPU
time for it to be sweeped trough, thus degrading performance.
2. Physically connected to the CPU... the CPU needs to have enough pins
to address all of the memory.

In all aspects, you are looking at a large surface to arrange all the
pins, making for a large computer, which again is not in favor of my
ultimate computer idea.

6) For a few more billion, MS will customize it's WinXP or Vista to
run in your hardware (one copy only)

I can do that without a single extra $. Just need some weeks and good
inspiration.

7) Someone will make you a few solid-state HD (the size will depend on
how much you are willing to spend).

The benefit of which would be what exactly? A large amount of this kind
of electronics could get damaged far more easily than any hard disks I
currently own.

8) For a few millions; I'm sure someone will make you a wall of nice
flat LCD/Plasma screens for you super-dupper one of the kind PC.

Okay that's the first good idea there. A large displaying surface. But a
high-resolution projector is a much much better solution than an
assembly of individual screens.

That would cover all major components - the small stuff is up to you.

The question is it practical? NO

Then obviously, you have not built the ultimate computer!

and probably no one ever crazy enough to do it (even if money is no
question).

If I have that kind of money; I wouldn't spend it on PC at all; I
would rather travell around the world in my jet or yatch and play with
women instead of sitting infront of the PC.

Well if all of those people at Intel and AMD were thinking that way,
we'd all still be using 8086s. For about the same amount of money as you
write about above, I might add.

--
Primary function: Coprocessor
Secondary function: Cluster commander

http://www.thought-beacon.net

Pay once per lifetime webhosting:
http://farcomm-it.com/?ref=jsah

We are the paragon of humanity. You may worship us. From afar.

01010010 01100101 01110011 01101001 01100100 01100101 01101110 01110100
01000010 01000001 01010011 01001001 01000011

 

[Jure Sah]

Mxsmanic pravi:

I'd want ultra-fast static RAM as fast as processor cache so that the
processor never waits for anything in memory at all, and I'd want
gigabytes of it.

That's a contradiction. A large amount of any kind of cache memory,
means a longer time for the CPU to scan trough it all. It also means,
since you can only put this much chips in one spot, longer physical
connections to the cache memory, further decreasing performance.

Theoretically you could have the cache be verified in a parallel manner,
but that, with larger amounts of it causes an ever greater amount of
circuitry you have to stick in one spot... eventually to such an extent
it becomes physically impossible to do so.

You see, there is a reason other than price why modern L1 caches are so
tiny.

I'd want disk drives at least a hundred times faster
than anything currently available.

You'd have to look at chips then. Hard drives are made physically and
can only spin so and so fast. You can't just multiply the factor by a
100 whenever you choose to.

I'd want network connections at
least one thousand times faster than anything currently available.

And serving you data from what servers?

That should cover most of the bottlenecks. The last thing I'd worry
about would be CPU speed.

Yes, but only in the event that all the hardware works perfectly
independently of the CPU. Actually, in that event, you don't even need a
CPU.

Also, if I were running an OS with a GUI, I'd want a separate GPU and
graphics subsystem with all of the same characteristics.

That's pretty much what you got today... And all of the graphical data
pre-loaded into the graphics memory perhaps? You have not mentioned what
kind of standard you would use to communicate between this graphics part
and the rest of the computer.

Also, you have the problem of display, the display has to have a refresh
rate comparable to the frame rate of the graphics subsystem, otherwise
there is little point in having it. LCD crystals take some time to morph
between states, CRT monitors also need time to scan, projectors use the
same technology as LCD, etc. How would you solve that?


--
Primary function: Coprocessor
Secondary function: Cluster commander

http://www.thought-beacon.net

Pay once per lifetime webhosting:
http://farcomm-it.com/?ref=jsah

We are the paragon of humanity. You may worship us. From afar.

01010010 01100101 01110011 01101001 01100100 01100101 01101110 01110100
01000010 01000001 01010011 01001001 01000011

 

[Mxsmanic]

They are for some ops like imaging and copying and editing large files.

Yes, but few people do such things on their computers.

 

[Mxsmanic]

If you got enough cache, the prefetching and caching will zero out the
access time in most situations. What kind of cases are we talking about?

For random access to large databases, cache is essentially useless;
every access to the database will generate physical disk I/O.

 

[Mxsmanic]

That's a contradiction. A large amount of any kind of cache memory,
means a longer time for the CPU to scan trough it all.

The CPU doesn't have to scan it.

It also means,
since you can only put this much chips in one spot, longer physical
connections to the cache memory, further decreasing performance.

It seems to work for supercomputers.

Theoretically you could have the cache be verified in a parallel manner,
but that, with larger amounts of it causes an ever greater amount of
circuitry you have to stick in one spot... eventually to such an extent
it becomes physically impossible to do so.

Perhaps you misunderstand. I'm not suggesting an expansion of cache.
I'm suggesting having all RAM operate at the same speed as on-chip
cache, so that even accesses outside of cache complete just as quickly
as cache hits. Of course, that's extremely expensive, but it can sure
boost the speed of the system.

You'd have to look at chips then. Hard drives are made physically and
can only spin so and so fast. You can't just multiply the factor by a
100 whenever you choose to.

Yes, unfortunately. But chips that can hold 500 gigabytes are hard to
find, and non-volatile chip memory tends to have very asymmetrical
read/write times, with writing taking vastly longer than reading.

And serving you data from what servers?

Any servers. I want the entire Internet to operate at that speed.

Also, you have the problem of display, the display has to have a refresh
rate comparable to the frame rate of the graphics subsystem, otherwise
there is little point in having it.

The upper limit on frame and refresh rates is set mainly by human
physiology. Anything above 30 fps or so is fine, and that's easy to
achieve.

 

[Phil_12345]

Jure Sah wrote:
(e-mail address removed) pravi:

Multicore system? Windows currently has poor support for even 2 cores,
what about more.

Please see #5

And I doubt that with those few billions of $$$, you'd be able to pull
out of the sky better technologies than are currently developed by the
companies.

Where, from outer space (alien technologies). How long you're willing
to wait for better technologies. Current computer technologies are
mostly improvement from previous design.

I took lessons in digital circuits and believe me, getting together a
CPU with 64 bits of functions (18446744073709551616) and put it all in
one optimized chip is not easy.

Like the OP; money is no object.

You've missed PCI64, it has a higher persistent transfer rate than
PCI-Express. But why limit oneself to such standards, if you created
your own hardware, you could plug it directly into the motherboard with
no slot interface and pick a communication frequency yourself.

And yet, there is always the question, if it's so easy, why don't they
do it?

Will the average consummer willing to pay high-price for it ??? Is
there a standard for it? NO and NO.

But if you have THAT kind of money and willing to pay; then someone
would willing to customize it for you.

Oh please, don't talk like you were born yesterday. An ISA Creative
sound blaster can be connected to as many other copies of the card as
you can fit to a bus.

The problem of course, is having a CPU with as much communication
bandwidth (for all those graphic cards). There are some physical limits
to how fast you can relay a signal without degrading it and now amount
of money can prevent them.

SLI is old technologies (3dfx; even Quantium has 4 or more voodoo1 SLI
on a single video card). Please show me a video vendor that has more
than 2 video cards SLI right now. If it's sooo yesterday then how
come hardly any improvement now a day.

Perhaps we would best dedicate a nuclear power plant to the thing.
However, IMHO, a computer that uses as much electricity as a small city
is not much of an ultimate computer at all.


No problem. However all memory needs to be:
1. Fast enough... if you have a lot of memory, it takes a lot of CPU
time for it to be sweeped trough, thus degrading performance.
2. Physically connected to the CPU... the CPU needs to have enough pins
to address all of the memory.

In all aspects, you are looking at a large surface to arrange all the
pins, making for a large computer, which again is not in favor of my
ultimate computer idea.


I can do that without a single extra $. Just need some weeks and good
inspiration.


The benefit of which would be what exactly? A large amount of this kind
of electronics could get damaged far more easily than any hard disks I
currently own.


Okay that's the first good idea there. A large displaying surface. But a
high-resolution projector is a much much better solution than an
assembly of individual screens.


Then obviously, you have not built the ultimate computer!


Well if all of those people at Intel and AMD were thinking that way,
we'd all still be using 8086s. For about the same amount of money as you
write about above, I might add.

Well, the OP was hypothetically (sp) speaking "if you have all the
money in the world" -- This is not true for people like me and you.
Therefore; my response was also hypothethically (sp) and not up-to PC
specs or in depth of how things work. If the OP requires more
meaningful response then perhaps he/she should take this question to
MIT or INTEL or MS or .... to get a realistic answer.

 

[Rod Speed]

Rod Speed writes

Yes, but few people do such things on their computers.

Mindless silly stuff. Plenty image their systems and plenty
do basic editing of video files, most obviously with PVRs.

 

[John Doe]

John Doe pravi:

Oh yes... on mass m8!

That doesn't translate to English very well.

 

[John Doe]

Phil_12345 hotmail.com pravi:

Some of the local institutions have some supercomputers that
aren't currently being used. I once asked somebody who had access
to them if we could go and take them for a spin since they're not
being used anyway. I was disappointed to learn that supercomputers
are highly specialized and don't actually provide the kind of
processing power one could use in PC software.


Whatever you could afford for those billions of dollars, a
supercomputer is simply not the ultimate computer. But build a
supercomputer that is capable of running PC software very very
well, then you will see my point.

In case you didn't notice. This is a home built PC group.

 

[Jure Sah]

John Doe pravi:

Nonsense.

Not very talkative of you. If it's not true, then show me examples of
general-purpose machines that can boot any of the modern OSs from flash
drives.

--
Primary function: Coprocessor
Secondary function: Cluster commander

http://www.thought-beacon.net

Pay once per lifetime webhosting:
http://farcomm-it.com/?ref=jsah

We are the paragon of humanity. You may worship us. From afar.

01010010 01100101 01110011 01101001 01100100 01100101 01101110 01110100
01000010 01000001 01010011 01001001 01000011

 

[Jure Sah]

Mxsmanic pravi:

The CPU doesn't have to scan it.

Then the CPU will not be able to use it. All cache works by matching and
matching involves scanning trough it.

It seems to work for supercomputers.

Which supercomputers are we talking about?

Perhaps you misunderstand. I'm not suggesting an expansion of cache.
I'm suggesting having all RAM operate at the same speed as on-chip
cache, so that even accesses outside of cache complete just as quickly
as cache hits. Of course, that's extremely expensive, but it can sure
boost the speed of the system.

It is unlikely to be expensive. It is more likely impossible to implement.

Any servers. I want the entire Internet to operate at that speed.

But then it is no longer about your computer.

The upper limit on frame and refresh rates is set mainly by human
physiology. Anything above 30 fps or so is fine, and that's easy to
achieve.

I can notice a CRT flicker, when it's set to 60 Hz. And they say I'm
human too.

--
Primary function: Coprocessor
Secondary function: Cluster commander

http://www.thought-beacon.net

Pay once per lifetime webhosting:
http://farcomm-it.com/?ref=jsah

We are the paragon of humanity. You may worship us. From afar.

01010010 01100101 01110011 01101001 01100100 01100101 01101110 01110100
01000010 01000001 01010011 01001001 01000011

 

[Jure Sah]

John Doe pravi:

In case you didn't notice. This is a home built PC group.

As said, if you gave it some thought, you'd understand. But it seems
giving something thought is too much of a strain to you, so you rather
resort to criticizing, which is nothing but a waste of bandwidth.

--
Primary function: Coprocessor
Secondary function: Cluster commander

http://www.thought-beacon.net

Pay once per lifetime webhosting:
http://farcomm-it.com/?ref=jsah

We are the paragon of humanity. You may worship us. From afar.

01010010 01100101 01110011 01101001 01100100 01100101 01101110 01110100
01000010 01000001 01010011 01001001 01000011

 

[Jure Sah]

Mxsmanic pravi:

For random access to large databases, cache is essentially useless;
every access to the database will generate physical disk I/O.

Yes, but not in a time-dependent fashion. A write to the cache will not
make the application wait for the actual write to be made. Rather, the
data will go into the cache and the cache will independently write it to
the disk. Unless you are persistently writing in very different
locations of a very large database, a cache of 1 GB will pretty much
provide sufficient buffer space for your writing needs. The cache will
also allow for optimizations in writing (cache has 0 seek time, disk has
enormous seek time).


--
Primary function: Coprocessor
Secondary function: Cluster commander

http://www.thought-beacon.net

Pay once per lifetime webhosting:
http://farcomm-it.com/?ref=jsah

We are the paragon of humanity. You may worship us. From afar.

01010010 01100101 01110011 01101001 01100100 01100101 01101110 01110100
01000010 01000001 01010011 01001001 01000011

 

[Jure Sah]

Rod Speed pravi:

Mindless silly stuff. Plenty image their systems and plenty
do basic editing of video files, most obviously with PVRs.

Yes, honestly, almost every PC user out there uses his PC to view and
copy around video files. They are much bigger than the average file and
any improvements in disk data access speed have clearly visible effects.

My Linux computer has 512 MB RAM and Linux uses nearly all of it's free
memory for disk cache. I use free disk space on my Linux computer to
store movies and CD backups and in copying them back and forth, the
amount of cache Linux uses greatly improves performance (speed at which
files are downloaded from one computer and stored onto another) compared
to a pair of windows systems, where each uses only a fixed amount of
80 MB cache (+ 8 MB hardware disk cache).

A RAID0 array running on one of my Windows machines, with no additional
cache (other than 8+8 MB of hardware disk cache) also visibly improves
the performance, especially when transferring data: the array <-> LAN or
the array <-> RAM.

--
Primary function: Coprocessor
Secondary function: Cluster commander

http://www.thought-beacon.net

Pay once per lifetime webhosting:
http://farcomm-it.com/?ref=jsah

We are the paragon of humanity. You may worship us. From afar.

01010010 01100101 01110011 01101001 01100100 01100101 01101110 01110100
01000010 01000001 01010011 01001001 01000011

 

[Jure Sah]

(e-mail address removed) pravi:

Where, from outer space (alien technologies). How long you're willing
to wait for better technologies. Current computer technologies are
mostly improvement from previous design.

Money and time are two different things. I said if you had all the money
in the world, not if you had all the time in the world.

Like the OP; money is no object.

See above.

Will the average consummer willing to pay high-price for it ??? Is
there a standard for it? NO and NO.

Yes and Yes. See PowerPC.

But if you have THAT kind of money and willing to pay; then someone
would willing to customize it for you.

But developing entirely new technologies will not only cost you money,
there is always a question if it at all can be done and if yes, it might
take a very long time. You don't have that kind of time, you only got
the money.

SLI is old technologies (3dfx; even Quantium has 4 or more voodoo1 SLI
on a single video card). Please show me a video vendor that has more
than 2 video cards SLI right now. If it's sooo yesterday then how
come hardly any improvement now a day.

SLI is nothing but a term for "Scalable Link Interface" that nVidia
uses. I have here a Daimond 64 ISA-VLB that has a digital link interface
that can be used to bridge as many cards as you wish (there is an
in-plug and an out-plug). The link connects the card's memory and
rendering capability. Neither of which is particularly impressive.

Beside there is no particular reason to make the link digital, you can
use analogue VGA linkage to the same effect. The Voodoo 2 3Dfx graphics
accelerator is a standalone card that is only packed with memory and
independent 3D rendering chips (4 per card I think), it has an in-plug
and an out-plug for analog VGA, the 3D data is simply divited amongst
the chips and overlayed and thus you can chain up as many Voodoo 2s as
you have. I have tried 3 of them and it worked just fine. Voila.

And why is this something new now all of the sudden? Because the company
you're buying your graphic cards from is trying to tell you it came up
with something new and you've fallen for it.

Well, the OP was hypothetically (sp) speaking "if you have all the
money in the world" -- This is not true for people like me and you.
Therefore; my response was also hypothethically (sp) and not up-to PC
specs or in depth of how things work. If the OP requires more
meaningful response then perhaps he/she should take this question to
MIT or INTEL or MS or .... to get a realistic answer.

I am the OP and I have considered taking it to INTEL or AMD, but to tell
the truth all these companies are capable of saying is promoting their
own findings while never really having time to consider things a bit
more universally.

Beside, every computer science college in this world teaches people all
you need to know about hardware and software design that you need to
think this whole thing trough and give an intelligent answer. The only
reason why I don't do this myself is because I find doing this trough
discussion much less stressing.

Unfortunately, as you have come to conclude, I find it hard to find
other people who are willing to give these subjects as much thought as I
do. People prefer to believe hearsay over factually determining what
hardware performs better.

--
Primary function: Coprocessor
Secondary function: Cluster commander

http://www.thought-beacon.net

Pay once per lifetime webhosting:
http://farcomm-it.com/?ref=jsah

We are the paragon of humanity. You may worship us. From afar.

01010010 01100101 01110011 01101001 01100100 01100101 01101110 01110100
01000010 01000001 01010011 01001001 01000011

 

[Mxsmanic]

Yes, but not in a time-dependent fashion. A write to the cache will not
make the application wait for the actual write to be made. Rather, the
data will go into the cache and the cache will independently write it to
the disk.

For some types of processing, you have to wait for the physical write
to be completed before proceeding.

 

[Rod Speed]

For some types of processing, you have to wait for the physical write
to be completed before proceeding.

Not very many at all today.

 

[John Doe]

John Doe pravi:

Not very talkative of you. If it's not true, then show me examples
of general-purpose machines that can boot any of the modern OSs
from flash drives.

Create Bootable USB Flash Drive
http://ucsu.colorado.edu/~shaher/Bootable_USB.html

How to boot from a USB device
http://www.weethet.nl/english/hardware_bootfromusbstick.php

Make Your Thumb Drive Bootable
http://www.pcmag.com/article2/0,1895,1846793,00.asp

 

[Mxsmanic]

Not very many at all today.

Mission-critical OLTP systems require it, and they are very common.

 

[johns]

Rule of a good builder is, you only buy what you need.
If you can define your needs, then build to that degree,
and no more. That is your optimum machine. If, in the
future, your needs grow, then, if you knew what you
were doing in your original build, you will be able to
upgrade to meet your needs. This should be possible
over a period of about 4 years .. with the last year
requiring pricy upgrades that tell you that you need
to do the original job again. In choosing parts, you
need to look carefully at warranty and reliability.
Most of all, you need to ignore the hype from the
fan boys and store salesmen. Personally, I like
to watch the price curves, and buy just as the parts
start to go out of date, and are being driven off the
market by the sales bullshit. For example: The
Viewsonic VX2025WM 20 inch is $342 at Mwave.
It is adequately taller than the 19 inch and at 8msec
it MIGHT be a pretty good gamer. The Samsung
215tw 21 inch is teetering at $480. It just CAN'T
hold there much longer because I'm starting to
see the 2025wm in the game stores, and that
means the price of the 2025 is going to drop another
$50 or so. The Samsung 215 has got to drop
to stay competitive .. and I suspect it will drop
to around $380, and I will grab it. My present
system is mid 3 years and going strong, and
I'm slowly building up my next game box. I have
the BFG GF7900GTO in it, and I'm watching
Vista and Crysis to tell me where the market
might go. My present system runs CoD2 set
to dx7 .. runs it great. My new system runs
CoD2 maxed out on all settings in dx9c. That
is my development benchmark right now. I
have very good reliabilty using Gigabyte bundles
from Mwave .. good tech support and proven
quality. I have tested and reject the following
products: ASUS, NEC, LG, XFX ... because they
have a very high failure rate, and no tech
support at all. So far, in my new build, I have
about $1200 invested .. no monitor yet.
When the time comes to change to it, I
will sell my present system for about $600,
and I will be into a brand new game box that
is totally up to my present needs for about
$980. THAT is a perfect system.

johns