768-Bit

[RavenSEAL]

So, when do you think we'll start to see the ultimate end cards start to feature that kind of bus speed?

 

[Rifter]

When the GPU's can take advantage of that much memory speed.

 

[cusideabelincoln]

768-bit is slow, the GTX 295 already went past this with 896 bit.

 

[RavenSEAL]

768-bit is slow, the GTX 295 already went past this with 896 bit.

True, but you gotta remember that the GTX295 is a dual core card, so technically, it's 448 x 2, many people can't afford to dump that much, so let me reinstate my question a little bit, when do you think we'll start to see single core 768-bit cards.

Edit: Not to mention the fact that the card isn't truly effective compared to todays higher end GPUs.

 

[cusideabelincoln]

Well I don't think we'll see 768-bit from any single GPU for a long time. 4-5 years perhaps, maybe never thanks to new RAM technology? Only dual chips will get to 768-bit.

 

[Lonyo]

When it's necessary. Given that at the moment 384-bit max seems to be enough, and RAM speeds also increase, the increase in memory bus bits shouldn't need to be particularly significant.

AMD has actually gone back from a 512-bit bus to 256-bit because GDDR5 was fast enough, and NV managed to go from 512-bit for the GTX285 to 384 with the GTX480 while increasing the bandwidth from ~160 to ~177GB/s by going from GDDR3 to GDDR5.
It's cheaper (and better in the long run for a product) to go with faster RAM chips than a higher memory bus to increase bandwidth because RAM prices usually go down, but (I think) PCB costs don't tend to.

 

[taltamir]

AMD has actually gone back from a 512-bit bus to 256-bit because GDDR5 was fast enough, and NV managed to go from 512-bit for the GTX285 to 384 with the GTX480 while increasing the bandwidth from ~160 to ~177GB/s by going from GDDR3 to GDDR5.

right, and to explain why this happened. 512bit bus takes up twice the die space of 256bit bus.
if ram speed and GPU speed increase in tandem, there is no reason to increase the bus width. if the GPU speed increases much faster than the ram speed, then increasing the bus width can compensate and allows for greater bandwidth (to accommodate that extra GPU speed). when (individual) ram speed itself caught up, the bit width was safe to reduce.

 

[Lonyo]

right, and to explain why this happened. 512bit bus takes up twice the die space of 256bit bus.
if ram speed and GPU speed increase in tandem, there is no reason to increase the bus width. if the GPU speed increases much faster than the ram speed, then increasing the bus width can compensate and allows for greater bandwidth (to accommodate that extra GPU speed). when (individual) ram speed itself caught up, the bit width was safe to reduce.

Less layers on the PCB, which also reduces cost.
http://www.tomshardware.com/news/Nvidia-GeForce-GTX260,6984.html

http://www.anandtech.com/show/2376
And here's a link about the HD2900 vs HD3800 where the memory bus change from 512-bit to 256-bit saved them probably over 50 million transistors (just to back up your claim).

Of course, this all assumes the memory controller can handle high RAM speeds.

Given the 384-bit bus, we initially assumed NVIDIA was running in to even greater memory bus issues than AMD ran in to for the 5000 series, but as it turns out that’s not the case. When we asked NVIDIA about working with GDDR5, they told us that their biggest limitation wasn’t the bus like AMD but rather deficiencies in their own I/O controller, which in turn caused them to miss their targeted memory speeds. Unlike AMD who has been using GDDR5 for nearly 2 years, NVIDIA is still relatively new at using GDDR5 (their first product was the GT 240 late last year), so we can’t say we’re completely surprised here. If nothing else, this gives NVIDIA ample room to grow in the future if they can get a 384-bit memory bus up to the same speeds as AMD has gotten their 256-bit bus.

http://www.anandtech.com/show/2977/...-gtx-470-6-months-late-was-it-worth-the-wait-

 

[taltamir]

Less layers on the PCB, which also reduces cost.

excellent point as well, in addition to taking up more die space, they also take up more traces, which could lead to extra layers, which cost more money.

also thanks for all the other links and quotes. nice of you to do all the legwork 😛...

 

[Arkadrel]

So, when do you think we'll start to see the ultimate end cards start to feature that kind of bus speed?

maybe never?

*IF* vram people keep makeing faster ram, a small bus size => saved die space => cheaper chip => cheaper card/better profits.

You see there are 7Ghz chips out...

Amd is gonna use 5.5-6.0Gb/s vram chips at 256bit bus and get 176(5.5Ghz) - 192(6.0Ghz) GB/s memory bandwidth. Currently cards probably arnt that memory bandwidth limited, as long as the GPU doesnt need more memory bandwidth than faster vram Ghz rated chips can give, solong shall the bus width stay the same around 256bit.

next gen, maybe 7Ghz will be enough (224 GB/s with 256bit bus @7GHz vram)? or maybe by then 8Ghz vram are out (256 GB/s)?

One thing that could happend is prices of high speed vram could raise to crazy expensive rates, then cards would probably get bigger busses again.

Of course, this all assumes the memory controller can handle high RAM speeds.

This is another thing that could happend, both nvidia and amd could run into issues makeing small fast busses, and that again would force them to go up in size. Nvidia seems to have had a harder time with their memory busses, and have been useing bigger width busses to reach their needed memory bandwidth.

 

[sagrada]

There's actual physical design limitations above 512-bit for the PCB for a single GPU.. costs to go above it increase exponentially