NV40 Specs

[Cyberkiller6]

Here.

 

[TourGuide]

I'll believe it when I see it.

 

[Peter D]

Originally posted by: TourGuide
I'll believe it when I see it.

 

[Detselom]

now what would this "16MB DRAM " do?

 

[godmare]

Originally posted by: Detselom
now what would this "16MB DRAM " do?

"auf Chip integriert"

 

[Keysplayr]

Originally posted by: Detselom
now what would this "16MB DRAM " do?

Could it be that GPU's are evolving into CPU's? Do they need some sort of prefetch cache? Like 16MB L3 off die cache?
Interesting. Those are some hefty specs. I wonder if they will be adhered to.

Keys

 

[keitaro]

Looks good. But I'll take it with a grain of salt. I wouldn't be surprised if some of the engineers have already had a few ideas planned for the next core. But 16MB is truly a questionable part, and quite a big goal to have.

 

[Dack76]

And I'm spent....anybody got a napkin?

 

[Wurrmm]

Why is it always the Nvidia chips that people like to buzz about. Granted I am not a fanboy, and though I myself do enjoying seeing potential specs for future Nvidia cards, it would be nice to see some for other chip manufactures, like ATI, Matrox, or even S3.

 

[BoomAM]

I could see gfx chips start to include onchip cache in the future, just look at what its done for the GameCube, the ATI/ArtX Flipper chip has on chip 1T-SDRAM, and thats performing very well.

 

[Jeff7181]

Doesn't sound unreasonable... however, you would think if nVidia released info about a future GPU, they'd release it to a more reputable source. If that info was leaked, then who knows what could eventually happen with those "plans"

 

[Jeff7181]

Originally posted by: Wurrmm
Why is it always the Nvidia chips that people like to buzz about. Granted I am not a fanboy, and though I myself do enjoying seeing potential specs for future Nvidia cards, it would be nice to see some for other chip manufactures, like ATI, Matrox, or even S3.

So post some specs... go find some and we'll all look at them with you.

 

[Intelman07]

I won't worry about it untill they are under $200 or even put into production for that matter.

 

[godmare]

Originally posted by: Wurrmm
Why is it always the Nvidia chips that people like to buzz about. Granted I am not a fanboy, and though I myself do enjoying seeing potential specs for future Nvidia cards, it would be nice to see some for other chip manufactures, like ATI, Matrox, or even S3.

That's an interesting observation, and I must concur. The only thing I know about the R400 from ATi is that they plan to see what nVidia has to offer with the NV40 first; presumably they and Matrox just keep a tighter lid on things. nVidia evidently hasn't learned the negative effects of overhyping from the NV30.

 

[Deeko]

Does not look very plausable.

btw, no offense, but next time post a real, respectable source, not your own website.

 

[Cyberkiller6]

....

 

[PliotronX]

Originally posted by: keysplayr2003

Originally posted by: Detselom
now what would this "16MB DRAM " do?

Could it be that GPU's are evolving into CPU's? Do they need some sort of prefetch cache? Like 16MB L3 off die cache?
Interesting. Those are some hefty specs. I wonder if they will be adhered to.

Keys

Close, it's a logical enhancement because then you could store the framebuffer into this ondie cache which would make accesses to it much faster. Since the textures don't required this kind of access latency (mostly just bandwidth I think), we could keep the local off-die memory for textures and it would work just fine.

That's how I understand it, maybe one of our resident techies will stop by and give us a more technical explanation

 

[BFG10K]

Close, it's a logical enhancement because then you could store the framebuffer into this ondie cache which would make accesses to it much faster.

Unlikely as you're not going to store much in a 16 MB framebuffer. It'll probably simply be a much larger extension of current caches which often store geometry and pixel/vertex instructions.

 

[Mr.IncrediblyBored]

Originally posted by: BFG10K

Close, it's a logical enhancement because then you could store the framebuffer into this ondie cache which would make accesses to it much faster.

Unlikely as you're not going to store much in a 16 MB framebuffer. It'll probably simply be a much larger extension of current caches which often store geometry and pixel/vertex instructions.

Why not? A 1600x1200 front buffer at 32-bit color requires <8 MB of storage. There's plenty of space for a front + z buffer.

 

[BoomAM]

I think he means texture wise.
Although, that still shouldnt matter too much. Look at the Flipper Chip in the GameCube, is has 3mb of on-chip 1T-SRAM (2mb frame buffer, 1mb texture cache, 10.4gb/s texture read speed), and some of the textures some games, are amazingly high detailed, wo/ slowdown. The key, is the speed and bandwidth between the cache and the actual core of the card.

 

[BFG10K]

There's plenty of space for a front + z buffer.

You're quite right, if you continue to keep the front and back buffers separate you can store 1600 x 1200 in slightly less than 16 MB. Obviously you'd want the flip to happen as early as possible so that you were rendering to the DRAM as much as possible.

Even if you exceeded its size by going to the main VRAM it's still heaps faster than going out across AGP to the main memory. And if you can get virtual texturing to work in that situation then it'd be even faster.

Yes, I do see DRAM being very useful in the future. Hopefully they can get those costs down so we can have more of it.

 

[BFG10K]

I think he means texture wise.

The rest of the data is just fine to sit in the VRAM.

A DRAM framebuffer + VRAM data is still going to be much faster than a VRAM framebuffer + VRAM data.

 

[Cyberkiller6]

I agree

 

[Goi]

Could the 16MB DRAM be embedded ram(EDRAM) aal Bitboys Glaze3D?

 

[BoomAM]

Originally posted by: Goi
Could the 16MB DRAM be embedded ram(EDRAM) aal Bitboys Glaze3D?

I could be, but to save costs, i would think that S1-SRAM would be used. Its cheap, and fast.