NV48 Cancelled

[mczak]

Functional, but performance is way lower than it should be. Dig up some Far Cry 1.3 benchmarks that compare HDR rendering with the NV40 (GeForce 6800 variants) and the GeForce 6600.

FP16 blending runs at "half speed".

Another thing to consider is that HDR rendering is not -only- FP16 blending. FP16 texturing also costs memory bandwidth, texture filtering time, and other datapath resources, compared to plain 8-bit texturing.

Of course, your simplistic claim that it's somehow broken makes so much more sense.

I think what you're missing is that the 6600GT suffers much less of a performance hit than the 6800 cards. http://www.xbitlabs.com/articles/video/display/farcry13_7.html
In the 5 different HL2 benchmarks, a 6600GT actually beats a 6800Ultra half the time when HDR is enabled!

 

[Luminescent]

Perhaps they lowered the cost of FP blending in NV43; it doesn't necessarily mean NV40 was buggered.

Secondly, why are there asterisks next to the 6600GT entries in those Xbit Labs benchmarks?

 

[mczak]

Perhaps they lowered the cost of FP blending in NV43; it doesn't necessarily mean NV40 was buggered.

That's possible, but the cost really looks unreasonably high on NV40 (often exceeding a factor of 3 in resulting performance). And the NV43 is supposed to be a bug-fixed, cut-down NV40, it seems unlikely Nvidia would introduce major changes with that chip.

Secondly, why are there asterisks next to the 6600GT entries in those Xbit Labs benchmarks?

A good question. If I'd have to guess, I would assume it's to denote the 2nd test system (system 1 is agp, 2 pci-express), though the article doesn't say which cards were agp, which pci-e (it's safe to assume that the 6600GT was pci-express).

 

[DegustatoR]

A good question. If I'd have to guess, I would assume it's to denote the 2nd test system (system 1 is agp, 2 pci-express), though the article doesn't say which cards were agp, which pci-e (it's safe to assume that the 6600GT was pci-express).

http://www.xbitlabs.com/articles/video/display/farcry13_5.html

 

[SiBoy]

[quote="Chalnoth
3. Video processor (again, pretty much broken on the NV40).
[/quote]

What's wrong with the video processor? Not that I'm surprised given Nvidia and their video capability (or lack thereof).

 

[mczak]

A good question. If I'd have to guess, I would assume it's to denote the 2nd test system (system 1 is agp, 2 pci-express), though the article doesn't say which cards were agp, which pci-e (it's safe to assume that the 6600GT was pci-express).

http://www.xbitlabs.com/articles/video/display/farcry13_5.html

Well, yes, but where in that page does it say which graphic cards were pci-express and which were agp?

 

[JoshMST]

I guess from my own perspective I would be surprised if NVIDIA did in fact cancel NV48. I am still not certain that The Inq is right about this one, but due to the lack of any new word on the NV48 for the past month or so, I would not be surprised if NVIDIA did in fact cancel it. Nobody seems to be really complaining alot about the GeForce 6800's video processor, and the whole FP blending deal is also essentially a non-issue in most reviews that I see. If NVIDIA can save a couple of million by not finishing a chip, and still be able to address the market with current products, then I can see them cancelling this one and just working on the NV47.

This is disappointing, but then again the NV48 would have been essentially released for a short time until NV47 hits (Q1 and Q2 2005). Apparently the bean counters at NV didn't feel that the NV48 was going to give enough ROI, especially when the current NV45 is able to address the market they want. Apparently yield issues at IBM are not as bad as they used to be, and if NV can release NV45's running at 425 MHz, what is the point of releasing the NV48?

Oh well, I guess we shall see, but this cancellation story does seem to make sense in light of the lack of news about any new NV offerings this December (which I believe was the original timeline for the NV48 release).

 

[JoshMST]

Then again, perhaps NV48 is alive and kicking.... or is this the NV41? If these guys are to be believed, it has essentially the same dimensions as the NV40. So, why would a NV41 be the same size if it has only 3 quads and 5 VS? Seems like an awful lot of wasted space if that is the case. The game guys didn't try NVStrap on this number. Curiouser and curiouser.

http://www.gamepc.com/labs/view_content.asp?id=msi6800pcie&page=2

Credit Pete for this link!

 

[PatrickL]

nVidia is written nvidi a. Seems strange that a machine write so badly on a chip.

 

[KimB]

FP16 blending runs at "half speed". This means 4 pixels/clock on NV40, and 2 pixels/clock on NV43. There's isn't enough memory bandwidth otherwise.

"half speed" would mean 8 pixels per clock on the NV40. And it isn't a memory bandwidth issue. The GeForce 6600 beats out the GeForce 6800 Ultra in some benchmarks with FP blending enabled. If that doesn't point to something broken in the NV40, I don't know what does.

 

[Dave Baumann]

Its "FP16 Blending and filtering"; if NV40 can only blend 8 pixels per clock it can only blend 4 FP16 pixels.

 

[KimB]

What's wrong with the video processor? Not that I'm surprised given Nvidia and their video capability (or lack thereof).

Well, there's something wrong with it that prevents it from working in the NV40. These problems were fixed with the NV43.

 

[Xmas]

nVidia is written nvidi a. Seems strange that a machine write so badly on a chip.

Huh? The distance between I and D seems to be exactly the same as between I and A.
http://www.nvidia.com

 

[PatrickL]

Follow the link provided by josh and look at the chip.

 

[Ardrid]

Then again, perhaps NV48 is alive and kicking.... or is this the NV41? If these guys are to be believed, it has essentially the same dimensions as the NV40. So, why would a NV41 be the same size if it has only 3 quads and 5 VS? Seems like an awful lot of wasted space if that is the case. The game guys didn't try NVStrap on this number. Curiouser and curiouser.

http://www.gamepc.com/labs/view_content.asp?id=msi6800pcie&page=2

Credit Pete for this link!

That's NV41, or at least it should be. It's the new form of the vanilla 6800.

 

[JoshMST]

Well, it might just well be the NV41, but I am confused about a couple of things....

First off, why is it the same size as the NV40? If it only has 3 actual quads (and not four with one disabled) as well as 5 vertex shaders, then why is there no difference in size?

Also, I thought that the NV41 was made on TSMC's 110 nm process? Again, if this is NV41, shouldn't it be quite a bit smaller?

Not saying you are wrong, but that picture brings up many, many questions.

 

[Dave Baumann]

Never heard of NV41 in relation to TSMC 110nm, I'd always seen the rumours to be 130nm IBM.

 

[Ardrid]

You're right, it should be somewhat smaller. Maybe MSI is using NV40 because they haven't received the real NV41 yet. That doesn't get around the lack of a bridge chip though, which is something an NV40 PCIe based board would need. What we need is a side by side comparison to a 6800 GT/Ultra GPU. It might be smaller but it's hard to tell with the shot we've got. Also, NV41 would be using IBM's 130nm process.

 

[Tim]

"half speed" would mean 8 pixels per clock on the NV40. And it isn't a memory bandwidth issue. The GeForce 6600 beats out the GeForce 6800 Ultra in some benchmarks with FP blending enabled. If that doesn't point to something broken in the NV40, I don't know what does.

Firingsquad has some HDR benchmarks, they show the 6800GT outperforming the 6600GT with the margin expected:

http://firingsquad.com/hardware/nvidia_nforce_4_sli/page13.asp

More HDR benchmarks from firingsquad:

http://firingsquad.com/hardware/far_cry_1.3_midrange/

Edit:
The 6800 HDR performance problem might be a AGP-only problem.

 

[Dave Baumann]

NV40 is AGP, there is no packaged bridge or external bridge.