G450 - actual bandwidth?

[MadCat]

My question is simple: is the 64-bit memory bus a 64-bit DualBus (2*64bit * DDR ~ 256) or is it a single bus ? You see, if its a single 64-bit bus, youve got a POS gamingwise, but if its a DualBus, youve got the equivalant of a G400 MAX with runaway overclocking potential.

So which one is it? Anyone know?

------------------
This Signature Space FOR SALE / RENT

[Hunsow]

http://forums.murc.ws/ubb/Forum5/HTML/008481.html

[Masayver]

The bus from the local memory to the GPU is 64-bits wide and is single. DualBus has nothing do with local memory.

------------------
You have been Donimated

[Greebe]

Time delayed dual postings? Hmmm
I'm sure glad Masayver has the white paper on all this or I'd be so confused.

[Jord]

A temporal time anomaly flux thingy there, Greebe

Nothing to worry about

Jord.

[MadCat]

Ok then, its time I got some clarification on how the G400 works.

I have always assumed the G400 had one 128-bit read bus to the memory and 1 128-bit write bus to the memory (hence dualbus). This would tend to support benchmarks which show the G400 MAX approaching and exceeding GeForce SDR performance at higher resolutions. But from what youre saying, the G400 has a single 128-bit bus to memory, and has two 128-bit wide rendering pipelines?

Could use some feedback on this point.

------------------
This Signature Space FOR SALE / RENT

[Jord]

Here you go, MadCat...
http://www.matrox.com/mga/press_room...df/dualbus.pdf

Just as you said, one 128bit pipeline in, and one 128bit pipeline out

Jord.

[Masayver]

"I have always assumed the G400 had one 128-bit read bus to the memory and 1 128-bit write bus to the memory (hence dualbus). "

I think everyone assumed this. I did too, until I read that pdf, as well as some other hardware sites, a little more closely. It's the bi between the engine and the data buffers, not from the engine to the local memory. I've always thought the G400 performed like crap (and I'm not talking shiz, as I have a G400. I'm just saying how it is with my gaming experiences) and now I know why. Again, don't think I'm anti-Matrox. I'm anxious to see what their next video card will offer (after G450).

------------------
You have been Donimated

[MadCat]

Thanks for the link Jord, that really explains things. Now, even though the dual read/write busses are only between the engine and the data buffers, this still means higher theoretical bandwidth, given the fact that the 128-bit memory bus will be put to more optimal use with the data buffer queues. Just not a whole lot more. hence the slightly better performance at higher resolutions than other cards with similar SDR memory speeds.

Now heres my question: is there anything besides prohibitive cost preventing accelerator makers from creating a card with a 256-bit bus between engine and memory? I realize that the number of layers required for that many traces is unthinkable, but is there really any limitation as far as the memory technology goes?

------------------
This Signature Space FOR SALE / RENT

[franksch3]

Some hardware sites are calculating what the maximum practical fillrate of some chips is depending on their external memory bandwidth.

However these calculations don't take into account internal caches, internal bandwidth and how optimized the core itself is.

Using the 256 bit dualbus, the G400 is much more efficient in 32 bit color than a TNT2 with 128 bit internal bus, although their external bandwidth is the same.

Benchmarking my G400 at 150 MHz core and 200 MHz memory against a TNT2 Ultra at 150 MHz core and 200 MHz memory, we got the following fillrate numbers for 32 bit color, 32 bit z-buffer, 32 bit texture:

G400: 200 MTexels/s
TNT2: 150 MTexels/s

Just because of the more efficient internal bus the G400 scores much higher. My feeling is that although nVidia products are technically pretty good, they lack optimizations to make better use of the external bandwidth.

The GeForce2 is pretty cool, with pixel shaders, 800 MPixels/s THEORETICAL fillrate. However, because their chipsets always seem to be pretty external memory bandwidth dependant, the actual fillrate at 32 bit always lacks.

I can easilly see the ATI Radeon score higher in 32 bit than the Geforce2, even though it might only have a theoretical fillrate of 400 MPixels/s. I even expect Matrox next part to score better.

[Stringy]

franksch3
Interesting tests done on the 2 vid. cards..
I'm curious though, what did the 16bit look like??
and what Bench did you use?? 3Dmark??

Craig

[franksch3]

We used 3DMark 99 Pro (not MAX) as benchmark. We only tested fillrate (textures, framebuffer and z-buffer al 32 bit) to see how efficient the G400 was.

Personally I think that the Voodoo1/2/3 16 bit looks better since you can't see dithering because of the filters used. Both the TNT(2) and G400 don't use a filter at all. I have to say that the G400 wins in 16 bit quality from the TNT2, however I don't care at all for 16 bit. I went for the G400 because of it's great 32 bit performance.

If I still wanted 16 bit I would have kept with my Voodoo2. The last few months before I got my G400 I could notice the 16 bit dithering and banding and how ugly it was. 32 bit color is really an improvement and I can certainly notice the difference, no mather what others might say.

And at 32 bit the G400 is quality and speedwise the winner without doubt. The TNT2 and Voodoo3 owners just commented about the great picture quality of the G400.

The only small drawback is that nVidia's drivers are faster and better than Matrox drivers. So in low resolutions the G400 is a little bit slower than the TNT2 because of the less optimized drivers, not because it is slower in hardware.