AMD breaks Teraflop limit (with Firestream 9250 GPU PCIe express card).

AMD breaks Teraflop limit
In German:
http://www.heise.de/newsticker/AMD-k...00/from/atom10
The Firestream 9250 GPU PCIe express card is likely based on the new RV770 GPU,
as will be used in the Radeon 4800 series.
The 1 Teraflop is for single precision, i tis 200Gflops for double precion 64 bit.
The card uses 150W.

Jan Panteltje wrote:
AMD breaks Teraflop limit
In German:
http://www.heise.de/newsticker/AMD-k...00/from/atom10
The Firestream 9250 GPU PCIe express card is likely based on the new RV770 GPU,
as will be used in the Radeon 4800 series.
The 1 Teraflop is for single precision, i tis 200Gflops for double precion 64 bit.
The card uses 150W.

And the other interesting thing about this is that even though the ATI
and Nvidia cards are close in single-precision speed, around 1TF, in the
double-precision department, the ATI card owns the Nvidia card, it's
something like 200GF vs. 78GF.

Yousuf Khan

"Jan Panteltje" skrev i meddelelsen
...
AMD breaks Teraflop limit
In German:
http://www.heise.de/newsticker/AMD-k...00/from/atom10
The Firestream 9250 GPU PCIe express card is likely based on the new RV770
GPU,
as will be used in the Radeon 4800 series.
The 1 Teraflop is for single precision, i tis 200Gflops for double precion
64 bit.
The card uses 150W.


I wonder if it will be possible to get a passively cooled 9250?

--

/Pnic

http://groups.google.dk/group/dk.tes...7?dmode=source

On a sunny day (Fri, 20 Jun 2008 02:31:23 +0200) it happened "Peter
Nicolaysen" wrote in
:

"Jan Panteltje" skrev i meddelelsen
...
AMD breaks Teraflop limit
In German:
http://www.heise.de/newsticker/AMD-k...00/from/atom10
The Firestream 9250 GPU PCIe express card is likely based on the new RV770
GPU,
as will be used in the Radeon 4800 series.
The 1 Teraflop is for single precision, i tis 200Gflops for double precion
64 bit.
The card uses 150W.


I wonder if it will be possible to get a passively cooled 9250?

150 W of heat, that is a lot, imagine a 150W lightbulb.
I think that would be very difficult.

On Jun 20, 5:46*am, Jan Panteltje wrote:
On a sunny day (Fri, 20 Jun 2008 02:31:23 +0200) it happened "Peter
Nicolaysen" wrote in
:


I wonder if it will be possible to get a passively cooled 9250?

150 W of heat, that is a lot, imagine a 150W lightbulb.
I think that would be very difficult.

If Firestream makes a serious dent in the computational market, where
power consumption and computational density are important, we'll
probably see a low power and maybe even an ultra-low power version.
The future for certain kinds of computation is racks and racks of
something like this, where the power and space consumed by the
collective is more important the the muscle power of a single card.

Robert.

On Jun 16, 6:29*am, Jan Panteltje wrote:
AMD breaks Teraflop limit
In German:
*http://www.heise.de/newsticker/AMD-k...-Grenze--/meld....
The Firestream 9250 GPU PCIe express card is likely based on the new RV770 GPU,
as will be used in the Radeon 4800 series.
The 1 Teraflop is for single precision, i tis 200Gflops for double precion 64 bit.
The card uses 150W.

Is it possible to stick it into a Windows based PC and expect OS to
recognize
it as 10 CPUs (provided appropriate drivers), or are programs going to
access it
completely independently from OS say using OpenCL?

Regards,
Yevgen

On a sunny day (Tue, 24 Jun 2008 11:44:58 -0700 (PDT)) it happened Yevgen
Barsukov wrote in
:

On Jun 16, 6:29*am, Jan Panteltje wrote:
AMD breaks Teraflop limit
In German:
*http://www.heise.de/newsticker/AMD-k...renze--/meld.=
..
The Firestream 9250 GPU PCIe express card is likely based on the new RV77=
0 GPU,
as will be used in the Radeon 4800 series.
The 1 Teraflop is for single precision, i tis 200Gflops for double precio=
n 64 bit.
The card uses 150W.

Is it possible to stick it into a Windows based PC and expect OS to
recognize
it as 10 CPUs (provided appropriate drivers),

I do not think so.
As far as I know you will have to write the soft do do specific math on that card.
There is a SDK I think.

You would start with a mathematical problem, and look what parts best can be done by the GPU.
It is, at least at this time, not a PC boost card or such.


The Radeon HD4800 seems to beat the Nvideo competion though in many aspects,
here some article in German I just found:
http://www.heise.de/newsticker/Radeo...meldung/109922

Maybe you can run it through babelfish translation or some site like that.

Yevgen Barsukov wrote:
Is it possible to stick it into a Windows based PC and expect OS to
recognize
it as 10 CPUs (provided appropriate drivers), or are programs going to
access it
completely independently from OS say using OpenCL?


No, the GPU is not a general-purpose processor, even when you call it a
GPGPU (general-purpose GPU). It lacks some important functions that CPUs
have that are needed for basic running of a computer. The CPU is not
busy just doing maths all of the time, let alone floating-point maths.
The CPU runs many non-mathematical tasks, actually it mostly runs
non-mathematical tasks. You wouldn't even be able to boot your computer
up without a CPU.

The GPUs are optional processors specialized in floating-point maths,
which can be employed if a program needs it. The CPU has its own
floating-point facilities available. The CPU floating point is more
precise than GPU floating-point, but the GPU makes up for it in speed.
However, programs will usually not be aware what kind of GPU resources
are available to them, so they will usually use just the CPU
floating-point. So that's why it would be necessary to integrate a
standard interface into the OS to support GPU-based calculations in a
vendor-independent way. Initially, the vendor-independence will be
achieved by standards like OpenCL, but they will be plug-ins into the
OS. Eventually, OpenCL or some other similar standard will have to be
built right into the OS, non-optionally.

Yousuf Khan

On Jun 25, 5:45*am, Jan Panteltje wrote:
On a sunny day (Tue, 24 Jun 2008 11:44:58 -0700 (PDT)) it happened Yevgen
Barsukov wrote in


Is it possible to stick it into a Windows based PC and expect OS to
recognize
it as 10 CPUs (provided appropriate drivers),

I do not think so.
As far as I know you will have to write the soft do do specific math on that card.
There is a SDK I think.

http://ati.amd.com/technology/stream.../sdkdwnld.html

Robert.

On Jun 25, 4:45*am, Jan Panteltje wrote:
On a sunny day (Tue, 24 Jun 2008 11:44:58 -0700 (PDT)) it happened Yevgen
Barsukov wrote in
:

On Jun 16, 6:29*am, Jan Panteltje wrote:
AMD breaks Teraflop limit
In German:
*http://www.heise.de/newsticker/AMD-k...renze--/meld.=
..
The Firestream 9250 GPU PCIe express card is likely based on the new RV77=
0 GPU,
as will be used in the Radeon 4800 series.
The 1 Teraflop is for single precision, i tis 200Gflops for double precio=
n 64 bit.
The card uses 150W.

Is it possible to stick it into a Windows based PC and expect OS to
recognize
it as 10 CPUs (provided appropriate drivers),

I do not think so.
As far as I know you will have to write the soft do do specific math on that card.
There is a SDK I think.

You would start with a mathematical problem, and look what parts best can be done by the GPU.
It is, at least at this time, not a PC boost card or such.

The Radeon HD4800 seems to beat the Nvideo competion though in many aspects,
here some article in German I just found:
*http://www.heise.de/newsticker/Radeo...pft-Konkurrenz....

Maybe you can run it through babelfish translation or some site like that..

Thanks. I can read German. It is good info.

Do you think it would be fundamentally possible to Re-use graphic
cards GPU
for floating-point math done on non-graphic programs that run on the
PC, _without_
creating a special card based on same hardware such as FireStream?
E.g. while I am not running Crisis on my expensive graphic card,
could I use my graphic-card resources running Mathcad 10 times faster?
Also graphic cards being much more
mainstream compared to floating point accelerators, it looks like such
dual-purpose solution
would be much cheaper...

Regards,
Yevgen