PDA

View Full Version : Post your ideal GPU/VPU Specs


mikechai
02-20-03, 04:54 AM
** My Ideal NVIDIA's GPU Specs **

NVIDIA 2004 GPU Lineup

NV40
200-250 Million Transistors on 130-nm process
400-500 MHz Core
8MB Embedded DRAM (77M trans.)
1.4 GHz 256MB DDR-II Memory
256bit wide Memory bus
16 Flexible Pixel Rendering Pipelines
16 Vertex Shader Engines
204.8 GB/sec Bandwidth (eDRAM)
44.8 GB/sec Bandwidth (DDR-II)
16 Gigapixels per Second
DirectX 10
Price $399-499

NV41
100-125 Million Transistors on 130-nm process
400-500 MHz Core
800-1000 MHz 128MB DDR-II Memory
256bit Memory Bus
8 Flexible Pixel Rendering Pipelines
8 Vertex Shader Engines
31.3 GB/sec Bandwidth
8 Gigapixels per second
DirectX10
Price $199-299

nutball
02-20-03, 06:25 AM
1. Fast (as in like doo-doo off a shovel)
2. Cheap
3. Available now

:D

oh, and did I forget to mention

4. Quiet

mikechai
02-20-03, 09:28 AM
1. Fast (as in like doo-doo off a shovel)
2. Cheap
3. Available now



oh, and did I forget to mention

4. Quiet


Hey, I couldn't be more agree with you.:)

Kruno
02-20-03, 09:47 AM
1. Free for all
2. Infinite bandwidth/fillrate/processing/etc...
3. Not made by nVidia/Ati ;)

:D

Mod
02-20-03, 10:04 AM
Originally posted by K.I.L.E.R
1. Free for all
2. Infinite bandwidth/fillrate/processing/etc...
3. Not made by nVidia/Ati ;)

:D

:lol: I would post this !

MuFu
02-20-03, 10:12 AM
Originally posted by nutball
1. Fast (as in like doo-doo off a shovel)
2. Cheap
3. Available now

:D

oh, and did I forget to mention

4. Quiet

:)

MuFu.

Uttar
02-20-03, 10:39 AM
My dream GPU, considering I've still got to have at least some modest hopes of it being released before March 2004 ( which should, theorically, be the time I'll be upgrading after buying something really soon )

So, yeah, thus are my "optimistic expectations" for the NV40.

NV40:
- 130nm process, Low K, very mature ( enabling higher clocks because it's mature )
- 600 Mhz using cooling quieter than 60db
- 256-bit memory bus, 600Mhz GDDR-2 memory ( I'm betting GDDR-3 for the NV45 )
- 160M Transistors

"What? That's all? But nVidia is going to get crushed by ATI if that's the best they could get out by H1 2004! Even the R450 could kill it in absolutely every case..."

Well, let's get to the good stuff, shall we? And please don't say "That's WAY too much for such a transistor count!" before reading it all. After, please feel free to do so.

- Cine FX 2 - PS: PS3.0. + 2048 Instructions + 512 Temps
VS: VS3.0. + 2048 Instructions + 512 Temps

"What the?!" - Yeah... But let me explain.
Those limits are really kinda artificial. In practice, you've got to divide it all by the number of calculators requring shading work.
So, there's one huge cache for instructions, temps and other things. It's freaking huge. But it's for the *whole* chip.
Those numbers are actually assuming 16 Calculators. I don't know how many there really are even in the NV30, so you'd have to divide that by 16 and multiply it by the number of calculators.

Note that I'm talking of calculators here: there's no more difference between VS & PS. It's all a global bool of calculators, like in the rumored R400. That means those calculators are dynamically allocated: either VS or PS cannot be stalled because the other is slower than the other.

All those instructions & temps would be dynamically allocated, and not 128 Instructions / 32 Temps per calculator.

That brings me to my next point. Multiple calculators could share the same instruction space, if they use the same program.
It would have a small effect on performance ( 1 instruction lost, because there got to be a delay between the two ) without branching, and it would obviously might have a very bad effect on performance with dynamic branching.
That means that even though the 2048 limit is kinda artificial, it isn't truly artificial. You could have a 2000 instruction PS program, and every pixel using that PS program would simply have a small delay. The good news is that, in theory, a few instructions time lost isn't a big deal, it's a really small part of the 2000 instructions.
But then you couldn't have more than a 48 instructions VS... Ah well!

This brings me to my next point. That card should be able to work on multiple different pixel/vertex programs at once! This would dramatically reduce the importance of stalls, thus increasing performance.

Oh, and what about...

- A programmable primitive processor ( PPP ) , also using that pool of calculators.

So, basically, the PPP won't cost too many transistors. But it'll hurt performance when using it, compared to not using it. But I'd guess the quality difference might be worth it :)

Cool, eh? :cool:

But that's a VERY optimistic hope. It's nearly impossible we'd get such a thing with the NV40. We might get some nice memory-saving techniques I didn't talk about, however.


Uttar

borntosoul
02-21-03, 01:27 PM
well if the nv35 ends up being close to twice as fast as the nv30 (with aa & af turned up) then that will be good enough for me to upgrade . thats my dream gpu !

poursoul
02-21-03, 06:22 PM
:afraid: ............... what planet are you from uttar. how old are you in human and christian years?

Shinri Hikari
02-21-03, 07:39 PM
:wtf:

Nice specs uttar...:cool: :D

Shinri Hikari
02-21-03, 07:48 PM
Oh and add molecular circutry...:drooling: :bugeyes: :thinker: :cool2: probably not going to happen for a while, S***.

legion88
02-21-03, 10:09 PM
Okay, here's my ideal GPU lineup to date.

NVXX
8 TeraHz 8192gigabit 256TeraByte of embedded memory
8 TeraHz core
64 pipelines
4 TMU (for the older games, of course)
64 vertex shader engines, 8192 instructions (to one up Uttar, of course) and 2048 temps

And no transistors. The GPU is organic, grown in a lab. When it's outdated, you can eat it.

Uttar
02-22-03, 04:28 AM
Well, this thread isn't exactly about what's the most logical thing we'll get :) It's about optimistic hopes, ain't it?

Actually, what I've described got multiple rumored R400 concepts, such as sharing PS & VS units.
I've just gone a step further - so that might be more like the R500 or NV50, and not NV40.
But I'm pretty sure that's the type of direction the industry will get to.

And no, I didn't talk about TBDR. It just seems too optimistic... Ah well, would be nice anyway :)


Uttar

GlowStick
02-22-03, 03:25 PM
My GPU/VPU needs is

screw that, Im waiting for Star Trek Holodecks.

kyleb
02-23-03, 02:18 AM
Originally posted by poursoul
:afraid: ............... what planet are you from uttar. how old are you in human and christian years?


his profile (http://www.nvnews.net/vbulletin/member.php?s=&action=getinfo&userid=402) answers the second question anyway. ;)

Uttar
02-23-03, 06:37 AM
Originally posted by kyleb
his profile (http://www.nvnews.net/vbulletin/member.php?s=&action=getinfo&userid=402) answers the second question anyway. ;)

Yeah, that information is correct.
But I've seen people who can't even calculate age correctly based on a birthdate - are YOU able to? :P


Uttar

kyleb
02-23-03, 06:46 AM
lol, ya and i don't knock you because your 14 either, i have ment many a man with at least twice your years and well under half your brains. ;)


i just figured i would point out were poursoul could find the answer to his question. :angel2: