128 mb really helps high res + FSAA?

[LardArse]

I have a GF4 Ti4200 128 mb but it comes with slower rams (4ns). The 64 mb cards have faster rams and are cheaper. I feel like I made a mistake buying the 128 mb card from all the benchmarks I see. However, I do play games at very high res + FSAA. For Morrowind, I use 1024 x 768 FSAA 4Xs and for Jedi Knight, I use 1600 x 1200 FSAA 4xs. I would like to know if the extra rams really help in my situations because I don't have a 64 mb card to compare it to.
Another question, I can enable 64 tap Anisotropic filtering on D3D with a 3rd party program but I've heard AF for D3D is broken. So there's no way to get AF working on D3D?

 

[CarstenS]

Hi there,

I do not think that opting for the 128MB-model was a mistake. Just have a look at the more modern Benchmarks and Games like codecreatures and comanche4, they indeed show quite an improvement with 128MB of Video RAM, and my guess is, that is going to be the trend with more and more engines to come.

 

[pcchen]

It's simple. Actually, you can't use 1600x1200 with 4X FSAA with only 64MB video RAM
However, considering GF4 Ti 4200 can't run very fast at 1600x1200 4X FSAA for many games, 128MB of slightly slower video memory is a not-so-good decision to me.

There is no forced anisotropic filtering option for D3D in current drivers. However, you still can use other tools such as Rivatuner to do it. It is working for me.

 

[LardArse]

Hi there. Does it mean that even with the 128 mb card, 4Xs FSAA is disabled at 1600 x 1200?

 

[CarstenS]

Nope, it's just quite slow, but available.

 

[Teasy]

It's simple. Actually, you can't use 1600x1200 with 4X FSAA with only 64MB video RAM

Not with a standard IMR you can't no. But with Z-buffer compression (which Geforce 4 has) it takes up 51.3mb (at worst) and 43.3mb (at best) for the Z-buffer and framebuffer, which would certainly make it possible. On a Tile based Renderer 1600x1200x32 4xFSAA would be easy with only 64mb of ram, it would take only 7.3mb.

Unless you mean that the Geforce 4 64mb has 4xFSAA forcibly disabled at 1600x1200?

Also I'm not saying 128mb of ram doesn't help, because in newer games with lots of textures on an IMR at high res and with 4xFSAA 128mb certainly can help, in a big way, over 64mb.

 

[MPI]

Well even 1024x768x32 w/4X takes up 40.5MB for the framebuffer. That leaves just 23.5MB for textures and vertex buffers on a 64MB card.

Running a 64MB card in 1024x768x32 w/4X is actually pretty much like running a 32MB card in plain 1024x768x32. And a 32MB card is bound to run into texture thrashing today. Remember, when the benchmarks run lower compared to 128MB cards, it's just not 20% slower overall, it's just that the average fps. becomes lower, because in parts where the texture memory overflows = stutter. Stutter = BAD!

 

[Hellbinder]

Well, unless you force Texture compression.. then that 32mb of left over will handle a lot more at 4-1/6-1 ratio.

this is where the new Nv30 and Perhaps the R300 will make such a difference in FSAA. If they indeed use some form of Frame Buffer Compression scheme.

 

[pcchen]

Not with a standard IMR you can't no. But with Z-buffer compression (which Geforce 4 has) it takes up 51.3mb (at worst) and 43.3mb (at best) for the Z-buffer and framebuffer, which would certainly make it possible. On a Tile based Renderer 1600x1200x32 4xFSAA would be easy with only 64mb of ram, it would take only 7.3mb.

Unless you mean that the Geforce 4 64mb has 4xFSAA forcibly disabled at 1600x1200?

Since he was talking about GF4, I assumed IMR.

I don't know how you get the 51.3MB number. Z buffer compression does not help at all, since 4 bytes are still reserved for each subpixel (the compression rate is not guaranteed, and GF4 does not support 3 bytes per pixel for Z buffer). Therefore, the back buffer and the Z buffer need 60,000KB memory. The front buffer needs additional 7,500KB memory, thus the memory requirement exceeds 64MB (= 65,536KB).

 

[Teasy]

Since he was talking about GF4, I assumed IMR

Ok, I read it as a broad statement about all vid cards instead of specific to Geforce 4.. my mistake.

I don't know how you get the 51.3MB number. Z buffer compression does not help at all, since 4 bytes are still reserved for each subpixel (the compression rate is not guaranteed, and GF4 does not support 3 bytes per pixel for Z buffer).

I realise that the Z compression ratio is not guarenteed to be as high as advertised. But I used a worst case senario of only 2x compression, or does Z-buffer compression sometimes not work at all?

BTW, I didn't know that memory storage was actually reserved in advance with the assumption of no Z-buffer compression. I just assumed that the card would use as much memory as it needed for the Z-buffer (depending on what sort of compression it attained) and the rest would be free for something else. Is this deffinately not the case? So in the end Z-buffer compression only every helps bandwidth and does not help to save memory space?

 

[MPI]

]Well, unless you force Texture compression.. then that 32mb of left over will handle a lot more at 4-1/6-1 ratio.

this is where the new Nv30 and Perhaps the R300 will make such a difference in FSAA. If they indeed use some form of Frame Buffer Compression scheme.

"Forcing" texture compression isn't possible on GF series, and I'm not sure if that's optimal anyway. It's better if the developer decides what should be compressed and what not to. In any case there certainly _is_ a reason to make 128M RAM an option. BTW, even with compression 32M video ram is not nearly enough for JK2. I've tried, had to go down to second to highest texture resolution and 16-bit textures.

Like you said, hopefully they have better more efficient ways in next-gen boards. I'm not really an FSAA fan just yet, seeing how wasteful and inefficient present methods are, and with that I include all except maybe Matrox's. It's the step in the right direction in any case, even if it didn't reach all the way.

 

[pcchen]

I realise that the Z compression ratio is not guarenteed to be as high as advertised. But I used a worst case senario of only 2x compression, or does Z-buffer compression sometimes not work at all?

BTW, I didn't know that memory storage was actually reserved in advance with the assumption of no Z-buffer compression. I just assumed that the card would use as much memory as it needed for the Z-buffer (depending on what sort of compression it attained) and the rest would be free for something else. Is this deffinately not the case? So in the end Z-buffer compression only every helps bandwidth and does not help to save memory space?

Hmm... I don't know about the implementation detail of current Z compression schemes in GF4 or R200. However, I don't think they will have a very complex dynamic allocation system for Z buffer. That's too complex and perhaps even inefficient.

And of course, there are always some cases that make a lossless compression scheme fail. How frequently it happens in practical situations is another issue, though.

I tried 4XS on my GF4 Ti 4200 64MB, and it is not working at 1600x1200. So I think it is safe to say that GF4 with only 64MB can't support 4X FSAA at 1600x1200 32bpp.

 

[nAo]

Hmm... I don't know about the implementation detail of current Z compression schemes in GF4 or R200. However, I don't think they will have a very complex dynamic allocation system for Z buffer. That's too complex and perhaps even inefficient.

In fact on gf3/4 zbuffer compression just saves bandwith but not memory.
A full zbuffer is allocated all the time.

ciao,
Marco