Fast memorychips vs 512bit bus

[Dutch Guy]

Much of the price of a new videocard is because superfast expensive memorychips are used, but why don't they use a 512bit memorybus instead as that would instantly double the memory bandwith?

Or am I missing the obvious?

 

[Spit]

It also costs a fortune to design a circuit board with 512 tracks as opposed to the 256 bit ones. Thats why I think.

 

[Dave2150]

I remember when ATI made the great move of using a 256bit bus for their R300 chip (9700pro), NVIDIA made the big mistake of using the older 128bit bus, hence the reason behind ATI's performance crown for a long time.

NVIDIA said somewhere that they thought at the time a 256bit bus would be too expensive to use in retail cards, whilst ATI ran the risk of using it.

If a 256bit bus was so expensive then, I imagine its very expensive to design a 512bit board now, for the same reasons as then. Its got to happen sooner or later though.

 

[mmj_uk]

...and it requires a hell of a lot of extra transistors.

 

[Dutch Guy]

...and it requires a hell of a lot of extra transistors.

Good point, I knew I was missing something obvious, that combined with the investment.

But what I did notice is that a lot of videocard are bandwith limited in situations, this for instance:http://www.beyond3d.com/reviews/bfg/7800gtoc/index.php?p=13

The maximum performance increases we see here are around 9%, which mirrors the bandwidth improvement from the memory overclock.

Indicating that a 7800GT is bandwith limited so a doubling of memory bandwith would possibly mean a lot better performance

 

[JaFFa]

Doubtlessly one of the companies will do this to regain the crown. Any guess which it will be though. You would think Nv would have leatrnt from last time and be pushing for it. But then again ATi might be in a better position financially right now to take a hit since they seem to be powering 2 of the next gen consoles.

Heh, where will it end ?

 

[m3csl2004]

mem bandwisth matters a lot on cards - i should know, ive had no end of trouble with the memory on my 9800pro (now at 265, when stock is 330 (not a typo)

but it is so much better than the equivilent nvidia card, and id still have it over a x700/6600 because of the very fact that its 256bit rather than 128


aint there something about 512bit bus between the memory chips themselves?

 

[Cyber-Mav]

hmm, 512bit bus would be good stuff im sure they would make a saving using slower memory, e.g 1ghz ram on a 512 bit bus would be equivalent to 2ghz ram on a 256bit bus.

then it all goes down to wether the savings in cash over the design and implementation of the 512bit bus outweigh the cost of using faster memory on 256bit bus vs slower ram on 512 bit bus.

 

[mdjmcnally]

Isn't the memory controller on the X1900's 512bit but only connected externally at 256bits for the memory chips.

 

[megatron]

Isn't the memory controller on the X1900's 512bit but only connected externally at 256bits for the memory chips.

Ive heard the same.

 

[Dutch Guy]

Isn't the memory controller on the X1900's 512bit but only connected externally at 256bits for the memory chips.

Yes, after I saw yout rpost I went back to a X1800 review explaining that here: http://www.anandtech.com/video/showdoc.aspx?i=2552&p=4

But I will look at it again a bit later as it is way too difficult early in the morning

For quite some time, the high end in graphics memory architecture has been a straight forward 256-bit bus divided into four 64-bit channels on the GPU. The biggest issues with scaling up this type of architecture are routing, packaging and clock speed. Routing 256 wires from the GPU to RAM is quite complex. Cards with large buses require printed circuit boards (PCBs) with more layers than a board with a smaller bus in order to compensate for the complexity.

 

[D.P.]

The trouble is the smaller the fab process the harder it is to wire up a 512bit bus.

Soemthing will have to be designed for the interconnect, and thats not likely anytime soon. Especially if NV and ATI are moving towards socketed GPUs the memory interface needs to be as simple as possible. + 512bit memory will add lots of latency issues.

The next stop is GDDR4 which should appear on the G80 and R600, this will allow some crazy speeds and double to tripple the bandwidth

 

[m3csl2004]

light ray interconnects between the vpu and vram