NV50 another monster GPU ?

[Guest]

Archived from groups: alt.comp.periphs.videocards.nvidia,comp.sys.ibm.pc.hardware.video (More info?)

the new NV40 aka GeForce 6800 is a beast. 16 pipelines, 6 Vertex
Shaders, 222 million transistors, 400 Mhz core (maybe higher later on)
on a 0.13 micron process.
we get 600 million vertices/sec - 6.4 gigapixels of fillrate, both
peak performance figures.

what will *NV50* bring in 18 or so months from now (fall 2005) ?

I was thinking that Nvidia will probably shatter the 1 billion
vertices/sec and 10 gigapixel fillrate barriers.

Nvidia seems to be roughly doubling the amount of transistors with
each new GPU generation.

NV10 ~23 million
NV20 ~57 million
NV30 ~125 million
NV40 ~222 million
NV50 ~??? probably over 400 million

(that's about right even if I got a few of those *slightly* wrong)

going from NV30's 125M transistors to NV40's 222M transistors was done
on the same micron process size, 13 nm or 0.13 micron, AFAIK. quite
an acomplishment.

NV50 will no doubt use a smaller process, either 11 nm / 0.11 micron,
or 90 nm / 0.90 micron. so 400~500 million transistors should be
little or no problem for Nvidia's engineers and IBM's (or TSMC's)
fabs.

I'm not sure if Nvidia is going to go to 32 pixel pipelines with NV50
or what. It might be quite a departure from NV40. then again maybe
not.

Nvidia will almost certainly double the amount of vertex/polygon
processing from NV40 which has 6 Vertex Shaders @ 400 Mhz giving 600M
verts/sec peak.
safe to say NV50 will be over 1 billion verts/sec however they achive
that.

as for new features, other than Shader Model 4.0+ I have no idea what
is going to be in NV50. things like embedded memory would be a nice
change, but that would be un-nVidia-like

 

[Guest]

Archived from groups: alt.comp.periphs.videocards.nvidia,comp.sys.ibm.pc.hardware.video (More info?)

On 15 Apr 2004 20:56:37 -0700, radeonr420@yahoo.com (R420) wrote:

>the new NV40 aka GeForce 6800 is a beast. 16 pipelines, 6 Vertex
>Shaders, 222 million transistors, 400 Mhz core (maybe higher later on)
>on a 0.13 micron process.
>we get 600 million vertices/sec - 6.4 gigapixels of fillrate, both
>peak performance figures.
>
>what will *NV50* bring in 18 or so months from now (fall 2005) ?
>
>I was thinking that Nvidia will probably shatter the 1 billion
>vertices/sec and 10 gigapixel fillrate barriers.
>
>Nvidia seems to be roughly doubling the amount of transistors with
>each new GPU generation.
>
>NV10 ~23 million
>NV20 ~57 million
>NV30 ~125 million
>NV40 ~222 million
>NV50 ~??? probably over 400 million
>
>(that's about right even if I got a few of those *slightly* wrong)
>
>going from NV30's 125M transistors to NV40's 222M transistors was done
>on the same micron process size, 13 nm or 0.13 micron, AFAIK. quite
>an acomplishment.
>
>NV50 will no doubt use a smaller process, either 11 nm / 0.11 micron,
>or 90 nm / 0.90 micron. so 400~500 million transistors should be
>little or no problem for Nvidia's engineers and IBM's (or TSMC's)
>fabs.
>

nVidia's days with TSMC may be over, except for existing chip
production. IBM is working on a .065u process jointly with AMD,
whjo may be funding some of the development.
No doubt nVidia is taking a great interest in proceedings in this
area, and is unlikely to jump ship from IBM. In fact, it may be a very

good long-term reason why nVidia has gone the IBM route with the
silicon for the 5700 and the 6800.

IBM must also be as pleased as punch with the reaction to the 6800.
Excellent publicity for their processes. The largest signal processing

chip ever produced and commercially available in a product sold to
the retail public.

I'm sure that IBM have produced custom silicon for their own computers
of equal or greater size but not for sale to the public.

John Lewis


>I'm not sure if Nvidia is going to go to 32 pixel pipelines with NV50
>or what. It might be quite a departure from NV40. then again maybe
>not.
>
>Nvidia will almost certainly double the amount of vertex/polygon
>processing from NV40 which has 6 Vertex Shaders @ 400 Mhz giving 600M
>verts/sec peak.
>safe to say NV50 will be over 1 billion verts/sec however they achive
>that.
>
>as for new features, other than Shader Model 4.0+ I have no idea what
>is going to be in NV50. things like embedded memory would be a nice
>change, but that would be un-nVidia-like