Why the GeForce FX doesn't have a 256-bit memory bus

[creedamd]

http://www.theinquirer.net/?article=7642

Use at your own risk, inquirer is questionable.

Any flaming will be reported to God himself.

[Sazar]

I think they are pretty accurate...

if you go to samsungs website you will see that their ddr II memory is listed as being 128bit :)

this has been discussed in another thread I believe... but I cant be bothered to do a search right now :)

[StealthHawk]

then why not just use "only" 800MHz DDR and couple that with a 256bit bus. it would have yielded more bandwidth :rolleyes:

i mean, the r9700 supposedly supports both ddr and ddrII in its memory controller, would it really be that hard for nvidia to do it too?

[Uttar]

Actually, DDR-2 originally made a lot of sense: if you had a higher memory clock than GPU clock.
The original goal of nVidia is supposed to have been 400Mhz core clock for the NV30.
Thus, they could have used DDR-I if they wanted a 1:1 ratio.

And thus, they wanted to have 500Mhz ( 1Ghz effective ) memory with a 400Mhz core. Which would probably indeed have given even cheaper FSAA & Aniso ( in % only, of course ) ...
Nice idea, too bad there was no 625Mhz DDR2 for a 500Mhz core :P


Uttar

[Kruno]

Originally posted by StealthHawk
then why not just use "only" 800MHz DDR and couple that with a 256bit bus. it would have yielded more bandwidth :rolleyes:

i mean, the r9700 supposedly supports both ddr and ddrII in its memory controller, would it really be that hard for nvidia to do it too?

Yes, theoretically 25GB/s bandwidth compared to their current 16GB/s. :)


Uttar, you have kind of contradicted yourself. :)

[Uttar]

Originally posted by K.I.L.E.R
Yes, theoretically 25GB/s bandwidth compared to their current 16GB/s. :)


Uttar, you have kind of contradicted yourself. :)

Err, no I didn't. I don't understand what you were talking about... You do understand you were quoting Stealth, right?

To understand what I said, however, you've got to take particuliar care in seeing when I talk about DDR-I and when I talk about DDR-II ( or sometimes I call it DDR-2, or DDR2, I just can't decide myself :) )


Uttar

[rwolf]

I think the Nvidia engineers goofed. I don't think they were expecting the R300 to perform as well as it does.

All the plans for the memory type and bandwidth are probably arranged 8-10 months in advance or more. I would imagine Samsung need sufficient lead time.

[Smokey]

Some info in this article (http://www.sudhian.com/showdocs.cfm?aid=315&pid=1169)

Its a GF-FX article over on the old Viahardware.com now sudhian.com
The author shot off an email to someone who knows a bit about Enhanced Memory Systems, and this is the response he got.

“There is no difference between DDR1 and DDR-II in terms of how wide the bus can be. The only problem is on the application / implementation level, that is, if the bus is running at what, 1000 MHz, signal skew between the different I/O lines will be extremely difficult to control. Again, that is mostly a board level problem, now a general issue with DDR-II. The GeForce FX uses a 10 layer PCB already because of the signal integrity issues…”

Hope that adds something :p

[Sazar]

Originally posted by Smokey
Some info in this article (http://www.sudhian.com/showdocs.cfm?aid=315&pid=1169)

Its a GF-FX article over on the old Viahardware.com now sudhian.com
The author shot off an email to someone who knows a bit about Enhanced Memory Systems, and this is the response he got.



Hope that adds something :p

the gf FX uses a 12 layer pcb... unless they went back to 10 layer and I haven't kept up to speed on it :)

concerning using the 2 techs... I think sudhian may well be right BUT @ the moment AFAIK samsung only makes 128bit gddr2 memory modules... per their website last I checked (a few weeks back)... hence it would be logical for nvidia to go with a 128bit bus...

concerning the signal integrity issues... remmeber that the original specs for the gf FX and the design presumably were carried out to support the initial set clocks of 400/400... going to 500/500 is undoubtedly going to cause its own problems... which can be quite clearly seen...

for this reason I was also saying in another thread that I believe the demonstration with ddrII (which presumably was ALSO samsung) by ati probably involved an r300 board setup so that it was 128bit... again this is MY assumption... if hellbinder or someone else can clarify as to the specs of the card used and if it was a fully compliant 256bit memory bus card running @ regular spec speed of 325/310.. I will defer to their know-how... am just making educated assumptions here (take the word educated with a pinch of salt :D )

btw.. this is for the bandwidth people.. the r300 based 9700pro has a theoretical bandwidth of 19.2 gb/s or something like that... v/s 16 gb/s for the gf FX 5800 ultra... the non-ultra is around the 12gb/s mark...

[Lezmaka]

Originally posted by Sazar
concerning using the 2 techs... I think sudhian may well be right BUT @ the moment AFAIK samsung only makes 128bit gddr2 memory modules... per their website last I checked (a few weeks back)... hence it would be logical for nvidia to go with a 128bit bus...

The ram used really doesn't have anything to do with the bus width. The ram used on th GFFX Ultra is not 128bit, but 128Mbits, as in 128 megabits. 128 megabits = 16 megabytes and since there are 8 chips total, means theres 128 megabytes of ram. If Samsung doesn't have any 256Mbit chips, then all that means is we won't be seeing a GFFX with 256 megabytes of ram, unless they can somehow squeeze 16 chips on the card.

Here's a link (http://www.samsungelectronics.com/semiconductors/Graphics_Memory/GDDR-II_SDRAM/128M_bit/K4N26323AE/k4n26323ae.htm) to the RAM used on MSI's board.

The 4Mx32 GDDR2 is 134,217,728 bits of hyper synchronous data rate Dynamic RAM...134,217,728 bits = 16,777,216 bytes = 16,384 kilobytes = 16 megabytes per chip

[Sazar]

Originally posted by Lezmaka
The ram used really doesn't have anything to do with the bus width. The ram used on th GFFX Ultra is not 128bit, but 128Mbits, as in 128 megabits. 128 megabits = 16 megabytes and since there are 8 chips total, means theres 128 megabytes of ram. If Samsung doesn't have any 256Mbit chips, then all that means is we won't be seeing a GFFX with 256 megabytes of ram, unless they can somehow squeeze 16 chips on the card.

Here's a link (http://www.samsungelectronics.com/semiconductors/Graphics_Memory/GDDR-II_SDRAM/128M_bit/K4N26323AE/k4n26323ae.htm) to the RAM used on MSI's board.

134,217,728 bits = 16,777,216 bytes = 16,384 kilobytes = 16 megabytes per chip

k... I was under the impression that using 128bit memory with a bus that was of a different width would cause complications... not that it can't be done but was not feasible...

I suppose I was operating under the wrong assumption...

[rwolf]

The only reason Nvidia went with the 128-bit bus is because they used the same design they had on the Geforce 4, which was enhanced for DDR2 memory. I think they had their hands full trying to get their floating point scheduler to work.

I don't think it performed as well as they expected so they had to overclock the crap out of it. After overclocking the chip the memory bandwidth became a barrier.

[suburbanguy]

So the question now is, will NV35 use 256-bit bus and DDRII
(or GDDR3) and will it be out before ATI has R400?

Since NV30 is an old design now (it should have been out almost a year ago) hopefully that means NV35 is basicly finished and awaiting a ramp up? IMO Nvidia shouls have NV35 out on shelves in May-July before R400 hits in August-October.

[Lezmaka]

Definately wont be GDDR3, GDDR2 is already expensive enough. R400 will probably use GDDR2 or unless they have some way of saving even more bandwidth, regular DDR. GDDR3 would make the card way too expensive unless it's gonna pull a NV30.