Improvement in CUDA performance?

"Smarty" wrote in message
...
Vincenzo Mercuri wrote:

Vincenzo Mercuri wrote:


Always distrust NVIDIA "new entries", wait them to revise

I meant, wait for them (nvidia folks) to revise...

Thanks for your advice and comments Vincenzo. I am not in a big hurry,
and will wait to see what nVidia offers next. The GTX 470 is expensive,
and I would very much like to buy a card which ONLY improves CUDA
performance since I do not play video games or use other features of
the GPU in the same way that a gaming enthusiast requires. Maybe there
is some way to get a big CUDA increase without spending a lot of money.

Thanks again!

Something to consider: Any change in the power and cooling
requirements. My computer room already gets hot enough
with three computers running most of the time, one with a
9800 GT.

I'm currently looking for power specs for the GTX 460 and
GTX 465, to see if I can replace the 9800 GT with one of
them without making the computer room even hotter and
requiring a power supply upgrade to match.

If not, I'll have to decide between a GT 240 or waiting for a
lower power version of the GTX 400 series.

Robert Miles

"Smarty" wrote in message
...
I have the opportunity to replace an older nVida GeForce 8800GT video
card with a new nVidia GTX470. Is there a website or other source where
I can predict what type of improvement I should expect in Cuda-based
processing. I am specifically interested in knowing how video rendering
programs like TMPGExpress, ProShow Gold, or other such software improve
with the substitution of the newer video card.

Thanks in advance for any opinions and suggestions!

See if this gives you enough information:

http://www.nvidia.com/object/graphic...s_buy_now.html

The Relative Compute Performance column appears to be
what you want.

Robert Miles

Robert Miles wrote:


"Smarty" wrote in message
...
I have the opportunity to replace an older nVida GeForce 8800GT
video card with a new nVidia GTX470. Is there a website or other
source where I can predict what type of improvement I should expect
in Cuda-based processing. I am specifically interested in knowing
how video rendering programs like TMPGExpress, ProShow Gold, or
other such software improve with the substitution of the newer
video card.

Thanks in advance for any opinions and suggestions!

See if this gives you enough information:

http://www.nvidia.com/object/graphic...s_buy_now.html

The Relative Compute Performance column appears to be
what you want.

Robert Miles

Thanks very much Robert for providing the link as well as your other
comments. The table does indeed allow a direct comparison, and I will
assume that most if not all CUDA software can exploit the extra cores
in the higher performance cards.

Heat and power supply capacity are two huge concerns for me as well,
and I just completed an air conditioning upgrade as well as a monitor
replacement specifically to cope with this problem. I was especially
impressed with how much cooler my new LED-based monitor runs with 28
watts of dissipated power compared to my prior LCD monitor consuming
110 watts, both recent vintage Samsung higher-end monitors.

The latest high performance video cards are quite power hungry, and my
Dell stock power supply is not going to handle the demand adequately. I
am assuming that any video card upgrade I make will demand a power
supply replacement as well, since the current supply is 450 watts and
has to handle a nearly full case of 3 hard drives, 2 optical drives,
full RAM sockets, other PCI boards, and a hungry Intel Extreme QX9650
at 3.1 GHz. My present card, the nVida 8800 GT, is not exceptionally
demanding as far as I can recall, but I need to compare the GTX 470 to
the 8800GT to see what the actual power demand difference is.

Thanks again for the link!

Smarty wrote:
Robert Miles wrote:

"Smarty" wrote in message
...
I have the opportunity to replace an older nVida GeForce 8800GT
video card with a new nVidia GTX470. Is there a website or other
source where I can predict what type of improvement I should expect
in Cuda-based processing. I am specifically interested in knowing
how video rendering programs like TMPGExpress, ProShow Gold, or
other such software improve with the substitution of the newer
video card.

Thanks in advance for any opinions and suggestions!

See if this gives you enough information:

http://www.nvidia.com/object/graphic...s_buy_now.html

The Relative Compute Performance column appears to be
what you want.

Robert Miles

Thanks very much Robert for providing the link as well as your other
comments. The table does indeed allow a direct comparison, and I will
assume that most if not all CUDA software can exploit the extra cores
in the higher performance cards.

Heat and power supply capacity are two huge concerns for me as well,
and I just completed an air conditioning upgrade as well as a monitor
replacement specifically to cope with this problem. I was especially
impressed with how much cooler my new LED-based monitor runs with 28
watts of dissipated power compared to my prior LCD monitor consuming
110 watts, both recent vintage Samsung higher-end monitors.

The latest high performance video cards are quite power hungry, and my
Dell stock power supply is not going to handle the demand adequately. I
am assuming that any video card upgrade I make will demand a power
supply replacement as well, since the current supply is 450 watts and
has to handle a nearly full case of 3 hard drives, 2 optical drives,
full RAM sockets, other PCI boards, and a hungry Intel Extreme QX9650
at 3.1 GHz. My present card, the nVida 8800 GT, is not exceptionally
demanding as far as I can recall, but I need to compare the GTX 470 to
the 8800GT to see what the actual power demand difference is.

Thanks again for the link!

8800 GT 85.7W measured

http://www.xbitlabs.com/articles/vid...t_5.html#sect0

GTX 470 196.6W measured

http://www.xbitlabs.com/articles/vid...0_6.html#sect0

http://www.xbitlabs.com/images/video...x400_power.png

*******

(These two shows the 465 and 460 for comparison.)

http://www.xbitlabs.com/images/video...x465_power.png

http://www.xbitlabs.com/images/video...w460_power.png

(You can compare their assets here.)

http://www.gpureview.com/show_cards....=632&card2=628

Measured values while running FurMark could give higher power numbers,
but the video card driver may have an option to detect FurMark and
underclock the card so the card or the voltage regulator on the
card are not damaged.

The 460 1GB is easier on power, but when speed is all that matters,
a little extra heat is besides the point.

Paul

Paul wrote:

Smarty wrote:
Robert Miles wrote:

"Smarty" wrote in message
...
I have the opportunity to replace an older nVida GeForce 8800GT
video card with a new nVidia GTX470. Is there a website or other
source where I can predict what type of improvement I should
expect in Cuda-based processing. I am specifically interested
in knowing how video rendering programs like TMPGExpress,
ProShow Gold, or other such software improve with the
substitution of the newer video card.

Thanks in advance for any opinions and suggestions!

See if this gives you enough information:

http://www.nvidia.com/object/graphic...s_buy_now.html

The Relative Compute Performance column appears to be
what you want.

Robert Miles

Thanks very much Robert for providing the link as well as your other
comments. The table does indeed allow a direct comparison, and I
will assume that most if not all CUDA software can exploit the
extra cores in the higher performance cards.

Heat and power supply capacity are two huge concerns for me as well,
and I just completed an air conditioning upgrade as well as a
monitor replacement specifically to cope with this problem. I was
especially impressed with how much cooler my new LED-based monitor
runs with 28 watts of dissipated power compared to my prior LCD
monitor consuming 110 watts, both recent vintage Samsung higher-end
monitors.

The latest high performance video cards are quite power hungry, and
my Dell stock power supply is not going to handle the demand
adequately. I am assuming that any video card upgrade I make will
demand a power supply replacement as well, since the current supply
is 450 watts and has to handle a nearly full case of 3 hard drives,
2 optical drives, full RAM sockets, other PCI boards, and a hungry
Intel Extreme QX9650 at 3.1 GHz. My present card, the nVida 8800
GT, is not exceptionally demanding as far as I can recall, but I
need to compare the GTX 470 to the 8800GT to see what the actual
power demand difference is.

Thanks again for the link!

8800 GT 85.7W measured

http://www.xbitlabs.com/articles/vid...bliss-8800gt_5
.html#sect0

GTX 470 196.6W measured

http://www.xbitlabs.com/articles/vid...gf-gtx400_6.ht
ml#sect0

http://www.xbitlabs.com/images/video...gtx400_power.p
ng

*******

(These two shows the 465 and 460 for comparison.)

http://www.xbitlabs.com/images/video...5/gtx465_power
.png

http://www.xbitlabs.com/images/video...glh/gw460_powe
r.png

(You can compare their assets here.)

http://www.gpureview.com/show_cards....=632&card2=628

Measured values while running FurMark could give higher power numbers,
but the video card driver may have an option to detect FurMark and
underclock the card so the card or the voltage regulator on the
card are not damaged.

The 460 1GB is easier on power, but when speed is all that matters,
a little extra heat is besides the point.

Paul


Thanks Paul for the citations and comments. I was initially encouraged
to find that the quiescent / idle power consumption of the 470 was
nearly the same as my current 8800GT, but the video cards apparently
ramp up their demand a lot under load, doubling or tripling their idle
current. Facing an increase of over 110 watts of dissipation between
the 8800 and 470 under load, and using a Dell stock power supply with a
450 watt rating, I would imagine that I am near, or at, or above the
appropriate load for this supply. Dell offered a larger supply when I
bought this XPS as an option, but I (foolishly) refused to pay the
premium. To the extent that Dell honestly and conservatively rates
their supplies, I may actually be able to squeak out enough power to
run this 470, but I presume I will do so at the expense of longevity to
all of the components inside the case, the power supply in particular.

I am faced with a thorny dilemma, since moving to a faster Nehalem or
Gulftown would cost me a lot more, forces me to go to an X58
motherboard which is incompatible with my (excellent) SpursEngine cell
coprocessor board, and thus loses in rendering performance the 2X gain
of the Spurs while adding maybe a 2X rendering improvement in the
i7-980X. The net benefit of a new motherboard / CPU for (let's say)
$2000 is thus a wash when it comes purely to rendering improvement.

Gaining CUDA speed, on the other hand, comes at a comparatively lower
cost, and could get me a 2X or greater gain with 'merely' a power
supply and GTX 470 purchase, for (let's guess) $600-$700. By retaining
my SpursEngine coprocessor card and bumping my CUDA speed up by 2X, I
think I have arrived at about the best quasi-optimal configuration for
rendering. Since rendering is the only activity where I truly need a
big boost in MIPS/GFLOPS, this seems like the way to go.

Kinda' unfortunate that the only 2 SpursEngine cards on the market, one
from Edius (Grass Valley) and the other from Leadtek are not X58
compatible. Many people using either of them are kinda' stuck when it
comes to upgrading their mobo or computer.

Smarty wrote:


Thanks once again Paul for your help and insights. I did a bit of
research a few months back on the Dell 'community' support forum to see
what others with the same model XPS420 I own have done to employ larger
demand / newer video cards. Although their approach appears to be
mostly empirical rather than analytical, they universally pulled out
the Dell standard supply and substituted a considerable higher capacity
Corsair or other supply. The standard size made the swap quite simple
apparently, and finding a supply with the same or similar connectors,
fan locations, etc. was not too hard to accomplish.

I could instrument my supply, and have current transformers and
ammeters and all the type of measurement devices I could assemble in my
50+ year career in ham radio, electrical engineering, etc., and would,
no doubt, get down to the very specific supply requirements, but I am
likely going to go for a big honker Corsair, Thermatake, or other brand
supply with a few hundred watts of extra capacity and just swap it. The
Dell forum has some specific recommendations as to which supplies are
the simplest to swap and I may just go with one of them.

I had never realized that the individual rails were constrained in such
a way that the aggregate power capacity of the supply may not be
available. Thinking about it for just a moment, it makes complete sense
that each rail is likely to have its own rectifier bridge, filters, EMI
chokes, transformer windings, etc., etc., etc. And of course regulators
for each voltage are separate discretes as well.

I have seen in many references the frequent comment that power supply
vendors often exaggerate their output power capacity, often grossly,
and I imagine that relying on manufacturer claims is absolutely fraught
with problems. Presumably the more established and reputable supply
vendors don't play games with their numbers, and genuinely deliver RMS
watts in a steady state, sustainable fashion versus some fanciful
'peak' watts delivered only in brief transient surges before the supply
crow-bars.....

There seems to be some consistent respect for Corsair as a vendor, but
I need to do some more research before spending the next $600-$700 on
the video card and power supply.

Again, many thanks Paul. Your knowledge and helpfulness is awesome!


I suspect the individual rail limits, could be shared resources. Like
the 3.3V and 5V could be using a common transformer, and so either rail
being loaded, results in the transformer heating up. (Or, say, it could be
a limit on what is driving the whole thing, from the primary side.)

Remember that Corsair, like Antec, relies on contract manufacturing, and
if you wanted to research them, you'd need to find a web site or a private
forum, where the actual source of the supply was noted. For example,
Antec at one time was using ChannelWell (and my CWT based Antec failed
with bad caps). Antec now is getting some of their supplies made by
Delta (not sure if this is the same Delta which was around 30 years ago).
Tracing the source of some of these units, can help you form a picture
of whether a supplier like this, has big differences between their low
end stuff and their high end stuff or not. They might not buy them all
from the same manufacturer.

I've seen some pretty **** poor examples of contracting. There is
one company (who shall remain nameless), consisting of like four
guys in an office somewhere. They buy supplies from overseas. Based
on the way they treated customers, they barely had any idea how the
power supply actually worked. So to get into the business, as Corsair
has done, doesn't necessarily require brains. You can be as dumb as
a post, and get into that business. And that (unnamed) company is still
retailing power supplies to this day. They probably don't even own a
Chroma tester, so they can test warranty returns or the like.
I don't know who they buy their supplies from. By comparison, the
staff at PCpowerandcooling (PCpower, now owned by OCZtechnology),
knew their products and how they worked, and were head and shoulders
above the four dumb guys. Good staff can make a big difference, when
you're trying to get satisfaction from one of these outfits.

Some companies do make their own supplies. Seasonic makes supplies for
themselves, and supplies power supplies under contract to other vendors.
They were the first company to offer 80% efficient supplies, which
are now a lot more popular.

Fortron/Sparkle make their own supplies. But they don't tend to do
"boutique" supplies, and they've only done a couple supplies up
around 600-700 watts that I can remember.

There is at least one guy on the newsgroups, who knows a lot about where
the supplies come from. I'd suggest the jonnyguru.com site as a source of
information, but it's pretty hard to search the site and get decent
info that way.

For its amusement value, I'll offer my favorite schematic of an ATX
power supply. This will give you some idea how a generic design works.
The output value is basically defined by the multi-output transformer
turns ratio. Some of the new supplies, aren't made this way, but this
is an excellent way to get some idea how they do it. The architecture
of this supply, is why the ATX standard has references to "crossloading"
on the outputs. It's because the supply lacks independent, tight,
closed loop regulation on each output. Regulation is sloppy on the
supplies, on purpose, because it allows the supply to be physically
smaller, and cheaper to make. If you look at pictures of
some of the ATX supplies that had like five separate transformers
inside, the chassis of the supply is quite long.

http://www.pavouk.org/hw/en_atxps.html

Paul

"Smarty" wrote in message
...
Robert Miles wrote:


"Smarty" wrote in message
...
I have the opportunity to replace an older nVida GeForce 8800GT
video card with a new nVidia GTX470. Is there a website or other
source where I can predict what type of improvement I should expect
in Cuda-based processing. I am specifically interested in knowing
how video rendering programs like TMPGExpress, ProShow Gold, or
other such software improve with the substitution of the newer
video card.

Thanks in advance for any opinions and suggestions!

See if this gives you enough information:

http://www.nvidia.com/object/graphic...s_buy_now.html

The Relative Compute Performance column appears to be
what you want.

Robert Miles

Thanks very much Robert for providing the link as well as your other
comments. The table does indeed allow a direct comparison, and I will
assume that most if not all CUDA software can exploit the extra cores
in the higher performance cards.

[snip]

Thanks again for the link!

You're welcome.

You may want to check if the software has been made able to
handle Nvidia's new Fermi chip designs before you order any
of the GTX 400 series boards - the GPUGRID BOINC project
I'm participating in had to produce a new program version in
order to start using any Fermi-based boards.

A site that may help you with the power supply calculations:

http://www.extreme.outervision.com/p...ulatorlite.jsp

I used to be an electronics engineer, before a stroke forced
me to retire early and with little remaining movement in
my left hand.

Robert Miles

Paul wrote:

Smarty wrote:


Thanks once again Paul for your help and insights. I did a bit of
research a few months back on the Dell 'community' support forum to
see what others with the same model XPS420 I own have done to
employ larger demand / newer video cards. Although their approach
appears to be mostly empirical rather than analytical, they
universally pulled out the Dell standard supply and substituted a
considerable higher capacity Corsair or other supply. The standard
size made the swap quite simple apparently, and finding a supply
with the same or similar connectors, fan locations, etc. was not
too hard to accomplish.

I could instrument my supply, and have current transformers and
ammeters and all the type of measurement devices I could assemble
in my 50+ year career in ham radio, electrical engineering, etc.,
and would, no doubt, get down to the very specific supply
requirements, but I am likely going to go for a big honker Corsair,
Thermatake, or other brand supply with a few hundred watts of extra
capacity and just swap it. The Dell forum has some specific
recommendations as to which supplies are the simplest to swap and I
may just go with one of them.

I had never realized that the individual rails were constrained in
such a way that the aggregate power capacity of the supply may not
be available. Thinking about it for just a moment, it makes
complete sense that each rail is likely to have its own rectifier
bridge, filters, EMI chokes, transformer windings, etc., etc., etc.
And of course regulators for each voltage are separate discretes as
well.

I have seen in many references the frequent comment that power
supply vendors often exaggerate their output power capacity, often
grossly, and I imagine that relying on manufacturer claims is
absolutely fraught with problems. Presumably the more established
and reputable supply vendors don't play games with their numbers,
and genuinely deliver RMS watts in a steady state, sustainable
fashion versus some fanciful 'peak' watts delivered only in brief
transient surges before the supply crow-bars.....

There seems to be some consistent respect for Corsair as a vendor,
but I need to do some more research before spending the next
$600-$700 on the video card and power supply.

Again, many thanks Paul. Your knowledge and helpfulness is awesome!


I suspect the individual rail limits, could be shared resources. Like
the 3.3V and 5V could be using a common transformer, and so either
rail being loaded, results in the transformer heating up. (Or, say,
it could be a limit on what is driving the whole thing, from the
primary side.)

Remember that Corsair, like Antec, relies on contract manufacturing,
and if you wanted to research them, you'd need to find a web site or
a private forum, where the actual source of the supply was noted. For
example, Antec at one time was using ChannelWell (and my CWT based
Antec failed with bad caps). Antec now is getting some of their
supplies made by Delta (not sure if this is the same Delta which was
around 30 years ago). Tracing the source of some of these units, can
help you form a picture of whether a supplier like this, has big
differences between their low end stuff and their high end stuff or
not. They might not buy them all from the same manufacturer.

I've seen some pretty **** poor examples of contracting. There is
one company (who shall remain nameless), consisting of like four
guys in an office somewhere. They buy supplies from overseas. Based
on the way they treated customers, they barely had any idea how the
power supply actually worked. So to get into the business, as Corsair
has done, doesn't necessarily require brains. You can be as dumb as
a post, and get into that business. And that (unnamed) company is
still retailing power supplies to this day. They probably don't even
own a Chroma tester, so they can test warranty returns or the like.
I don't know who they buy their supplies from. By comparison, the
staff at PCpowerandcooling (PCpower, now owned by OCZtechnology),
knew their products and how they worked, and were head and shoulders
above the four dumb guys. Good staff can make a big difference, when
you're trying to get satisfaction from one of these outfits.

Some companies do make their own supplies. Seasonic makes supplies for
themselves, and supplies power supplies under contract to other
vendors. They were the first company to offer 80% efficient
supplies, which are now a lot more popular.

Fortron/Sparkle make their own supplies. But they don't tend to do
"boutique" supplies, and they've only done a couple supplies up
around 600-700 watts that I can remember.

There is at least one guy on the newsgroups, who knows a lot about
where the supplies come from. I'd suggest the jonnyguru.com site as a
source of information, but it's pretty hard to search the site and
get decent info that way.

For its amusement value, I'll offer my favorite schematic of an ATX
power supply. This will give you some idea how a generic design works.
The output value is basically defined by the multi-output transformer
turns ratio. Some of the new supplies, aren't made this way, but this
is an excellent way to get some idea how they do it. The architecture
of this supply, is why the ATX standard has references to
"crossloading" on the outputs. It's because the supply lacks
independent, tight, closed loop regulation on each output. Regulation
is sloppy on the supplies, on purpose, because it allows the supply
to be physically smaller, and cheaper to make. If you look at
pictures of some of the ATX supplies that had like five separate
transformers inside, the chassis of the supply is quite long.

http://www.pavouk.org/hw/en_atxps.html

Paul

Last things first...... The amusement value of the schematic and
narrative is very entertaining, and it does reinforce your description
of the design philosophy.

Regarding Corsair, I have no particular preference for brands given my
utter lack of experience with DIY PCs of any recent vintage. (The last
time I built a computer was a Southwest Technical Products in the
1970's!) but I have seen both Corsair and PCPower mentioned often. I
was entirely unaware of how the designs, fabrications, and support
concepts now exist except to assume that everything is off-shore, race
to the bottom, minimalistic designs to drive down cost. I have to think
that the high end gaming machines from Alienware and others must be the
exceptions with some quality parts, conservative designs, etc., but
never saw one in person.

I am going to watch Newegg for some "shell shockers", Black Friday, or
other deal for a highly rated supply with 600 or more watts, and assume
it will do the job. Hope my case temperature doesn't rise, but I assume
that my fan noise will unavoidably increase given that the GTX470 fan
and other fans will be stressed way beyond my present meager 8800GT.

I was surprised to hear your comment that 80% efficiency supplies are
apparently what typify the better designs. I would have guessed that
switching supplies with higher efficiencies were more common. 20% seems
like a lot of wasted energy / heat.

Smarty wrote:


I was surprised to hear your comment that 80% efficiency supplies are
apparently what typify the better designs. I would have guessed that
switching supplies with higher efficiencies were more common. 20% seems
like a lot of wasted energy / heat.


If you think 80% is wasteful, there are still supplies out there
at the standard 68% level, and those kick out a lot of heat.
So 80% is an improvement.

Have a look at a review, to see how good some of these are getting.

http://www.jonnyguru.com/modules.php...print&reid=153

Paul

Paul wrote:

Smarty wrote:


I was surprised to hear your comment that 80% efficiency supplies
are apparently what typify the better designs. I would have guessed
that switching supplies with higher efficiencies were more common.
20% seems like a lot of wasted energy / heat.


If you think 80% is wasteful, there are still supplies out there
at the standard 68% level, and those kick out a lot of heat.
So 80% is an improvement.

Have a look at a review, to see how good some of these are getting.

http://www.jonnyguru.com/modules.php...=print&reid=15
3

Paul

Thanks Paul. I have learned a lot here, and the Corsair review leaves
me extremely impressed with both the jonnyguru website and the specific
Corsair being reviewed. I will not be needing anything quite so
powerful, but I will now do some research on this and other websites to
see what a good choice would be for my 600-650 watt purchase.

Many thanks again!