NV40 Speculation... will it really fix the DX9 issue?

[TheTaz]

Ok... NFI's GPU Rumor Watch sez about the NV40:


In order of decreasing reliability:

- Supports FP32, FP16 and FX16 natively. Whether there is any performance difference between FP16 and FX16 is unknown, and whether there are any truly non-FP32 units is also unknown.

- 175M transistors, 600Mhz core clock, 1.5Ghz effective GDDR2 ( Samples already shipped to nV - 16 memory chips per board )

- Not taped-out yet, or if it did, tape-out failed. To tape-out sometime this month.

- 8 pipelines, Speculation: probably 8x2 and no 16 zixel trick ( not worth it with 4x+ AA, which is really a minimum with 48GB/s of bandwidth ) - Maybe such a bypass path for low-end models ( NV42/NV43 )

This means Christmas avaibility has IMO become absolutely out of the question. Best we can hope is for-developer documents for Comdex, or around then. And that's not a certainty.


Ok... I realize this is all RUMOR.... but I don't see FP24 in there... do you?

Supposing NFI is RIGHT... do you think FP16 will be good enough?

Taz

PS: A link for those who don't know about NFI GPU Rumor Watch:
http://www.notforidiots.com/GPURW.php

[lukar]

As long as they make new chip based on NV3x it will be new failure and the end of Nvidia.

NV40 have to be a totally new design, chip so on...

[mrsabidji]

Many interesting infos can also be found here :

http://www.notforidiots.com/forum/viewtopic.php?t=8

mrsabidji

[TheTaz]

Originally posted by lukar
As long as they make new chip based on NV3x it will be new failure and the end of Nvidia.

NV40 have to be a totally new design, chip so on...

I may be wrong... but my understanding is that the NV3X doesn't have FP16.... only FX16....but the rumor shows both in the NV40.

Everything I've read about the DX9 Fiasco relates to FP24.

So... I guess I'm saying... I'm confused why there is no FP24 in nVidia's 3rd attempt at a DX9 range of cards. ;)

Regards,

Taz

[bkswaney]

Originally posted by TheTaz
I may be wrong... but my understanding is that the NV3X doesn't have FP16.... only FX16....but the rumor shows both in the NV40.

Everything I've read about the DX9 Fiasco relates to FP24.

So... I guess I'm saying... I'm confused why there is no FP24 in nVidia's 3rd attempt at a DX9 range of cards. ;)

Regards,

Taz


The NV3X can do 16 and 32. 32 is way to slow on the NV3X.

NV blew it by not using 24. 24 is the DX9 min. :)

I think the NV3X can also do 12.
They r having to drop back and run 16 to keep up with ATI.
But 16 is below the min DX9 standard.

[Hellbinder]

my opinion is that this entire last year has been very Harsh for Nvidia. Shocking, eye opening, slap in the face etc.

I simply cannnot believe that they would let that happen to them again. Not like this. My guess is they are going to go all out, money is no object, absolutely insane as they can possible make the specs with current technology. These Guys need to hit a Home Run of 500' into the parking lot. A Doubble or Tripple is not going to cut it.

Here is where I peg the Nv40 at.

-8 Pixels per clock no exceptions. As crazy as it seems I will even bet 16 pixels in some cases.

- PURE FP32 throughout the entire card. Period and nothing else.

-Total paralellism between Texture ops and Shader ops. none of this losing the use of FP units for texture work.

-Enhanced FSAA. Something new that is going to be unique to them. With Higher Quality than they have ever put out before.

-High Performance AF Perhaps even with a hardware implimentation similar to ATi.

-500mhz Core

-600-700mhz Ram

Well there you have it. Thats my take. A card that will deliver all the goods that should have been delivered last November. With some enhancements to boot. This is a no compromize card. Peak IQ with out even an option to cut it down. More than Doubble the performance of the Nv35.

[TheTaz]

Originally posted by bkswaney
The NV3X can do 16 and 32. 32 is way to slow on the NV3X.

NV blew it by not using 24. 24 is the DX9 min. :)

I think the NV3X can also do 12.
They r having to drop back and run 16 to keep up with ATI.
But 16 is below the min DX9 standard.

Ok... then what the hell is the difference between FX16 and FP16?

And Why would nVidia still not conform to the DX9 Spec of FP24, with their 3rd attempt at a DX9 range of cards, when they are obviously getting pounded???

Taz

[bkswaney]

Originally posted by Hellbinder
my opinion is that this entire last year has been very Harsh for Nvidia. Shocking, eye opening, slap in the face etc.

I simply cannnot believe that they would let that happen to them again. Not like this. My guess is they are going to go all out, money is no object, absolutely insane as they can possible make the specs with current technology. These Guys need to hit a Home Run of 500' into the parking lot. A Doubble or Tripple is not going to cut it.

Here is where I peg the Nv40 at.

-8 Pixels per clock no exceptions. As crazy as it seems I will even bet 16 pixels in some cases.

- PURE FP32 throughout the entire card. Period and nothing else.

-Total paralellism between Texture ops and Shader ops. none of this losing the use of FP units for texture work.

-Enhanced FSAA. Something new that is going to be unique to them. With Higher Quality than they have ever put out before.

-High Performance AF Perhaps even with a hardware implimentation similar to ATi.

-500mhz Core

-600-700mhz Ram

Well there you have it. Thats my take. A card that will deliver all the goods that should have been delivered last November. With some enhancements to boot. This is a no compromize card. Peak IQ with out even an option to cut it down. More than Doubble the performance of the Nv35.


I agree HB... They r going to have to make a killer card next time.
You know ATI is. ;)

[reever2]

Ok... then what the hell is the difference between FX16 and FP16?

FX is integer and FP is floating point

[TheTaz]

Originally posted by Hellbinder

-8 Pixels per clock no exceptions. As crazy as it seems I will even bet 16 pixels in some cases.

- PURE FP32 throughout the entire card. Period and nothing else.

-Total paralellism between Texture ops and Shader ops. none of this losing the use of FP units for texture work.


So I take it you believe those 3 things would replace FP24, and then some? (DX9 performancewise)

Regards,

Taz

[mrsabidji]

Originally posted by TheTaz
Ok... then what the hell is the difference between FX16 and FP16?

And Why would nVidia still not conform to the DX9 Spec of FP24, with their 3rd attempt at a DX9 range of cards, when they are obviously getting pounded???

Taz

If nv40 can achieve doing FP32 faster than ATI/DX9's FP24 there will be no use for it at all. You never know. I know some will tell me nividia engineers are morons and stuffs. I would say the same about those at ATI back in Radeon7xxx time and see how wrong I was...;)

mrsabidji

[TheTaz]

Originally posted by reever2
FX is integer and FP is floating point

That was my guess.
However, I'm still not clear on wether NV3X only has FX16 or both FX16 and FP16. ;)

Regards,

Taz

[TheTaz]

Originally posted by mrsabidji
If nv40 can achieve doing FP32 faster than ATI/DX9's FP24 there will be no use for it at all.

Ok... that makes sense.

Between what you've said and what Hellbinder said... I feel more educated, now.

Thanks! :)

Taz

[mrsabidji]

Originally posted by TheTaz
That was my guess.
However, I'm still not clear on wether NV3X only has FX16 or both FX16 and FP16. ;)

Regards,

Taz

I think nv30 has FX12, FP16 and FP32 whereas nv35 has only FP16 and FP32. Correct me if I'm wrong.

mrsabidji

[Hellbinder]

So I take it you believe those 3 things would replace FP24, and then some? (DX9 performancewise)

Regards,

Taz

Ok FP16, FP24, FP32 etc has absolutely nothing to do with the speed of the card. ATi could just as easily made a card with FP24 that was slow as ice melting in the Arctic.

Its about how said technology is implimented. Whether or not the engineers put the harware under the hood to actually handle the load. The Correct number of Registers, the caches needed, the proper layout, adequate "parts" to accomplish the Job requested in a single clock cycle.

It is in this regard that ATi Succeeded and Nvidia Failed miserably. ATi's card is an 800 pound FP24 crunching Gorilla. Nvidias are Chimpanzee bananna eating one sec, swinging from the ceiling the next, FP16/FP32 Zoo animals. If Nvidia had placed their Emphasis on FP16 or FP32 and sold the farm on it none of this would be happening to them today.

FP32 is better than FP24 in every way except Transistor count and Cost to impliment. When someone writes a DX9 shader they dont say FP24 or FP16 or FP32. The non tweaked code in its pure form simply chooses the Highest it can as the default. Thus FP24 for ATi and FP32 For Nvidia. Once Nvidia produces a card with the Needed internals to push FP32 then they have the upper hand virtually overnight against ATi's FP24.

If ATi sticks with FP24 with Loki and Nvidia pushes FP32 to the next level with Nv40. Its a Whole Different ball game. Then it simply comes down to who is the fastest and which one looks the best. The IQ Crown would go to FP32 100% of the time in a pure enviornment.

[The Baron]

What Hellbinder said, both times.

[Dirty]

its gonna fixz everything.

nvidia made a single mistake with fx (16/32 , rather than standard 24)

i dont think they will repeat that ever again. they learnt their lessons !

next nv cards r gonna kick ass

[mrsabidji]

They'd better...;)

mrsabidji

[TheTaz]

Excellent post to *un-cross* my eyes (and possibly a few others) to the situation, Hellbinder.

Thank you VERY much! :)

Regards,

Taz

[The Baron]

Despite what people say, ATI screwed up Big Time with the R200 (Radeon 8500). They recognized this, skipped the refresh, and went to R300. The rest is history. The chances that NVIDIA will have the same luck on their next-generation part are pretty high.

[Hellbinder]

Despite what people say, ATI screwed up Big Time with the R200 (Radeon 8500). They recognized this, skipped the refresh, and went to R300. The rest is history. The chances that NVIDIA will have the same luck on their next-generation part are pretty high.

The Radeon 8500 is a good hardware design. Its problem was it was released 6 months to late. Introduce that card around the same time as the GF3.. Whole different ball game. The one weak area with it is its memory controler was not nearly as efficient as Nvidias. Still it would totally have outperformed the Origional GF3 had it been introduced in a timely manner.

[The Baron]

Originally posted by Hellbinder
The Radeon 8500 is a good hardware design. Its problem was it was released 6 months to late. Introduce that card around the same time as the GF3.. Whole different ball game. The one weak area with it is its memory controler was not nearly as efficient as Nvidias. Still it would totally have outperformed the Origional GF3 had it been introduced in a timely manner.
and if the launch drivers didn't suck

That's what killed it--by the time the drivers were decent, the GF4 was smashing it.

[mrsabidji]

Originally posted by The Baron

quote:
--------------------------------------------------------------------------------
Originally posted by Hellbinder
The Radeon 8500 is a good hardware design. Its problem was it was released 6 months to late. Introduce that card around the same time as the GF3.. Whole different ball game. The one weak area with it is its memory controler was not nearly as efficient as Nvidias. Still it would totally have outperformed the Origional GF3 had it been introduced in a timely manner.
--------------------------------------------------------------------------------


and if the launch drivers didn't suck

That's what killed it--by the time the drivers were decent, the GF4 was smashing it.

It strangely resembles the nv3x story in almost every way, except nvidia didn't skip the refresh, which made their failure look even worse.

mrsabidji

[The Baron]

Originally posted by mrsabidji
It strangely resembles the nv3x story in almost every way, except nvidia didn't skip the refresh, which made their failure look even worse.

mrsabidji
Well, yeah. But, NV35 is an iffy refresh. It changed PS slightly and added a 256-bit memory bus, and that's all it added. It dropped the DDR2 and the Dustbuster. So, it's a very minor refresh at best.

[Red Dog]

Originally posted by Hellbinder
my opinion is that this entire last year has been very Harsh for Nvidia. Shocking, eye opening, slap in the face etc.

I simply cannnot believe that they would let that happen to them again. Not like this. My guess is they are going to go all out, money is no object, absolutely insane as they can possible make the specs with current technology. These Guys need to hit a Home Run of 500' into the parking lot. A Doubble or Tripple is not going to cut it.



Imo, Nvidia will pull out all stops this round. That should be a good thing.