NV40 a 16-pipe MONSTER - Too late for ATI to respond

[NV55]

According to an article and several posts I read this morning, the
NV40 is a 16 pipeline design with 205~210 million transistors--up from
the 8 pipeline configuration / 175 million transistors of eariler
reports.

but *I* don't believe this info yet.

NV40 as a 16-pipe architecture will *probably* end up as a 8x2. much
like NV30 was thought of, at first, as an 8-pipe architecture. then
later it was discovered NV30 is a 4x2 (that's 4 pipes, 2 texture
units per pipe).


now, here's the NV40 article in question:

-------------------------------------------------------------------------------
Nvidia NV40 specifications confirmed

Too late for ATI to respond


By Paul Dutton: Thursday 26 February 2004, 14:52

NVIDIA REALLY needed to pull the cat out of the bag in the next round
of its increasingly epic battle with ATI Technologies and from where
we're sitting we'd be more than surprised if the next round isn't,
now, a foregone conclusion.
US sources close to Nvidia have confirmed that their next generation
GPU will feature a full sixteen pipelines – not as previously
speculated an 8x2 arrangement - and this is reflected in the increased
transistor count of circa 205-210 million, up from a previously
speculated 175 million.

In very recent times ATI have been extremely confident in saying that
their R420 - set to launch not long after, Nvidia show their hand –
would, immaterial of architecture, thoroughly outgun the NV40.

Indeed today, Richard Huddy - ATI's European Developer Relations
Manager maintained that their expectation was for ATi's R420 to be,
faster, on balance, across a suite of 10 common games than Nvidia
NV40.

However we think that ATI's earlier confidence was based on internal
intelligence that Nvidia's NV40 feature set would be limited to an
architecture built on 175 million transistors and that Nvidia would
deliver, as anticipated, on the first day of CeBIT in March.

Nvidia's launch date for NV40 seems to have shifted backwards, but
even a month or so is not going to give ATi enough time for a hardware
response.

Knowing what we know now about NV40 having a full sixteen pipelines,
we sensed that despite some fighting talk, and indeed some intelligent
and rationalised arguments, ATi are now hoping, at best for a close
fight…

NV40 0wnage of ATi R420 coming soon to a store near you?

We sure think so… µ

 

[GTX_SlotCar]

Nvidia's launch date for NV40 seems to have shifted backwards, but

even a month or so is not going to give ATi enough time for a hardware
response.

In the past, they've pushed backwards and backwards until the missed the
product cycle. Let's see how they do this time. Looks like good news for
nVidia owners, though. It means another release of new Det. drivers that are
very fast (and unstable) to bring up the new card's scores. I wonder how
they'll work in my kids' Ti4400.

Gary

 

[John Lewis]

According to an article and several posts I read this morning, the
NV40 is a 16 pipeline design with 205~210 million transistors--up from
the 8 pipeline configuration / 175 million transistors of eariler
reports.

but *I* don't believe this info yet.

NV40 as a 16-pipe architecture will *probably* end up as a 8x2. much
like NV30 was thought of, at first, as an 8-pipe architecture. then
later it was discovered NV30 is a 4x2 (that's 4 pipes, 2 texture
units per pipe).

Either 175 million or 200 miliion transistors on even a
90nm process will present a huge cooling challenge.
Ask Intel and Prescott (only ~ 120 million transistors ).

It will be very interesting to see how this will all pan out.
nVidia is working actively on using IBM's latest
processes (and IBM is working with AMD also on
a new 65nm process) so it will be very interesting
to see what emerges in the GPU world before the end
of 2004. Heat management is becoming a very serious
issue for both CPUs and GPUs, particularly for the
GPUs which are now rapidly outpacing CPUs in terms
of both transistor count and active clock-edges.

John Lewis

 

[kevin getting]

Either 175 million or 200 miliion transistors on even a
90nm process will present a huge cooling challenge.
Ask Intel and Prescott (only ~ 120 million transistors ).

Intel's thermal issues stem from the actual design of Prescott based P4's.
90 nm technologies are not the issue, just ask IBM.

It will be very interesting to see how this will all pan out.
nVidia is working actively on using IBM's latest
processes (and IBM is working with AMD also on
a new 65nm process) so it will be very interesting
to see what emerges in the GPU world before the end
of 2004. Heat management is becoming a very serious
issue for both CPUs and GPUs, particularly for the
GPUs which are now rapidly outpacing CPUs in terms
of both transistor count and active clock-edges.

I don't expect GPU's to carry high Mhz ratings since they exploit
parallelism so well. It probably easier to double the number of pipelines
in the NV40 than it would be to double the clock speed. ATI and nVidia
are not affraid to launch a part with a lower Mhz rating than the previous
generation as long as the chips are more effecient. PCI Express hopefull
will address some of the heat and power issues.

 

[Derek Baker]

--
Derek

Either 175 million or 200 miliion transistors on even a
90nm process will present a huge cooling challenge.
Ask Intel and Prescott (only ~ 120 million transistors ).

The NV40 is still on 130nm. Imagine cooling that! Still they had plentry of
experience - mostly bad - cooling the NV30.

 

[papasurf]

As long as it doesn't have six fans, heat up the entire house, sound like a
chain saw and take up four PCI slots

 

[Dark Avenger]

*snip snip snip*
--------------------------------------------------------------------------------

http://www.theinquirer.net/?article=14373

The Inquirer? Mm, sorry Fox News told me that new Ati GPU was a 64
Pixel Pipeline with 6 TMU's per pipe, so nvidia is so slow in
comparison!

... the above just means that the source is pretty worthless and that
Fox News might even have more value!

 

[Courseyauto]

--
Derek

Either 175 million or 200 miliion transistors on even a
90nm process will present a huge cooling challenge.
Ask Intel and Prescott (only ~ 120 million transistors ).

The NV40 is still on 130nm. Imagine cooling that! Still they had plentry of
experience - mostly bad - cooling the NV30.

,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,
,,,,,,,,,,,,,,,,,,,

ATIknows all about bad experiencs,their drivers prove that...............

 

[Courseyauto]

As long as it doesn't have six fans, heat up the entire house, sound like a
chain saw and take up four PCI slots

According to an article and several posts I read this morning, the
NV40 is a 16 pipeline design with 205~210 million transistors--up from
the 8 pipeline configuration / 175 million transistors of eariler
reports.
<<<<<<<<<<<<<<<<<<<<>>>>>>>>>>>>>>>>>>>>>>>>>>

You mean like the 5900..........................

 

[Courseyauto]

The Inquirer? Mm, sorry Fox News told me that new Ati GPU was a 64
Pixel Pipeline with 6 TMU's per pipe, so nvidia is so slow in
comparison!

... the above just means that the source is pretty worthless and that
Fox News might even have more value!

<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<


Like most of the information in the newsgroups......................

 

[Dark Avenger]

The Inquirer? Mm, sorry Fox News told me that new Ati GPU was a 64
Pixel Pipeline with 6 TMU's per pipe, so nvidia is so slow in
comparison!

.. the above just means that the source is pretty worthless and that
Fox News might even have more value!

<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<


Like most of the information in the newsgroups......................

Yup, it's just waiting untill the first official benchmarks show up,
okay benchmarks can be cheated and thus should not be taken 100%
literal true, yet... they are a nice indication!

 

[Egil Solberg]

PCI Express hopefull will address some of the heat and
power issues.

I cannot see why that should have any influence.

 

[joe smith]

I cannot see why that should have any influence.

Maybe it will consume 0.1 watts less, who knows, if it takes less
transistors to implement the PCI E interface than the AGP one.

 

[John Lewis]

Intel's thermal issues stem from the actual design of Prescott based P4's.
90 nm technologies are not the issue, just ask IBM.

You are missing some facts. Intel's 90nm process has leakage
current problems that they are trying to remedy, which is why
the present Prescott performance is so unsatisfactory. Nothing to do
with the chip DESIGN. Intel may carry out some design tweaks
to get around process problems, but there is nothing in the chip
architecture that has caused the thermal problems. It was expected
that the power-increase due to the larger caches and pipelines
would be more than offset by the process shrink. It has not worked
out that way.... so far.

By the way, what high-volume publicly-available parts are IBM making
today on their 90nm process ?

I don't expect GPU's to carry high Mhz ratings since they exploit
parallelism so well. It probably easier to double the number of pipelines
in the NV40 than it would be to double the clock speed.

Elementary my dear Watson. doubling the pipelines will double the
power for that part of the architecture..........you will get a heat
problem one way or the other, if you want to boost the graphics
computational power by a factor of 2. Lowering the clock rate
will just lower that factor. Nothing new here. DSPs use massive
parallel-processing. Whether you process in parallel or serial, it
takes a given number of clock-edges within the chip per second
to get the job done and the total number of clock-edges executing per
second everywhere within the chip is roughly proportional to the power
consumed.. For a given functionality, the only white horse to come in
and save on power will be a process shrink, as long as the design
stays in silicon or Si-Ge; lower capacitance and smaller voltage
swings are the key power-saving ingredients.

ATI and nVidia
are not affraid to launch a part with a lower Mhz rating than the previous
generation as long as the chips are more effecient.

More efficient ! What do you mean ?

PCI Express hopefull
will address some of the heat and power issues.

PCI Express has nothing to do with the fundamental problem.
It may save some I/O power, but that is a trivial portion of
the total power consumed by a GPU.

John Lewis

 

[John Lewis]

*snip snip snip*
--------------------------------------------------------------------------------

The Inquirer? Mm, sorry Fox News told me that new Ati GPU was a 64
Pixel Pipeline with 6 TMU's per pipe, so nvidia is so slow in
comparison!

.. the above just means that the source is pretty worthless and that
Fox News might even have more value!

Well, we all know that......... but however it pans out, the next
generation GPUs will have to come up with clever approaches
to heat problems.

The nub of the problem:-

The silicon-design tools are getting so good that very
large designs can be accomplished in very short time spaces.
and is far outstripping the ability to shrink the processes within
that time-scale. In fact process-shrinks are coming
slower and slower with exponentially increasing expenses
for each iteration. Thus the silicon area will get larger, at
least in the near-term and the thermal-solution vendors
will make a lot of money.

John Lewis

 

[Darthy]

As long as it doesn't have six fans, heat up the entire house, sound like a
chain saw and take up four PCI slots

Hey... howabout they build in the sound of the fans eating up Dawn...
like as if she fell in a wood chiper... but not as fast...

The new Nvidia GFX6-GT 6800Ultra -"Its a SCREAM just turning her on!"