Can G70 use the new 700/800MHz GDDR3?

[Jawed]

I'm not at all proficient with the various kinds of DDR, so I'm just asking all this in general terms.

The new Samsung GDDR3 memory that is used on R5xx appears to have a different data format from previous generations of GDDR3 memory. 32-bit as opposed to 64-bit. Have I got that right?

How would NVidia connect this memory (to get 256/512MB of 700/800MHz) to G70? Is it actually possible?

Jawed

 

[trinibwoy]

Is it actually possible?

Jawed

Well I can't answer the more technical points. But they would have had to have known about this configuration long in advance and if they didn't build into G70 the capability to interface with the new memory setup it would have been really stupid on their part.

 

[rendezvous]

I'm not at all proficient with the various kinds of DDR, so I'm just asking all this in general terms.

The new Samsung GDDR3 memory that is used on R5xx appears to have a different data format from previous generations of GDDR3 memory. 32-bit as opposed to 64-bit. Have I got that right?

How would NVidia connect this memory (to get 256/512MB of 700/800MHz) to G70? Is it actually possible?

Jawed

Can you please be more elaborate on in what way they differ?

As far as I can see there aren't any major differences that necessarily would make it not work.

 

[_xxx_]

I'm not at all proficient with the various kinds of DDR, so I'm just asking all this in general terms.

The new Samsung GDDR3 memory that is used on R5xx appears to have a different data format from previous generations of GDDR3 memory. 32-bit as opposed to 64-bit. Have I got that right?

How would NVidia connect this memory (to get 256/512MB of 700/800MHz) to G70? Is it actually possible?

Jawed

You would take two 32-bit chips and run them as one 64-bit chip?

 

[Jawed]

Older memory chips appear to have 64-bit data interfaces and these new ones 32-bit.

I'm not sure if I've got the wrong end of the stick on this, though. And whether it actually makes a difference.

Jawed

 

[Jawed]

You would take two 32-bit chips and run them as one 64-bit chip?

Sounds obvious - which is why I made the thread to find out how it actually works, rather than just guessing.

Jawed

 

[rendezvous]

Older memory chips appear to have 64-bit data interfaces and these new ones 32-bit.

The K4J55323QF (256 Mbit, used on Quadro 4500 and i assume all G70 chips) chips are using 32 physical data connections, which is the same as the K4J52324QC (512 Mbit, which i ussume is used in the 520 since it's the only memory from samsung reaching that high klocks).

 

[Jawed]

Thanks.

So the 512Mbit chips could, literally, drop in as replacements for the eight 256MBit chips that currently appear on 7800GTX. Is that correct?

Jawed

 

[Rys]

Yes. And such a product is likely done and dusted already

 

[Tim Murray]

Yes. And such a product is likely done and dusted already

Oh, you mean for the 512MB G70 that will launch at 460Mhz two days after R520?

 

[Jawed]

Hey, why not call it the 7800Ultra, 525/1400.

Jawed

 

[Tim Murray]

Hey, why not call it the 7800Ultra, 525/1400.

Jawed

525 seems high. Between 460 and 480, though, seems Just Right.

 

[_xxx_]

Hey, why not call it the 7800Ultra, 525/1400.

Jawed

Yay!!! New thread!!!

 

[Geo]

525 seems high. Between 460 and 480, though, seems Just Right.

You're low, me thinks. The assumption is they'll double-slot the cooler too, maybe juice it a touch.

Is the feeling is they are "go" no matter what, or "only if it would catch R520XT"? And if they go that route, does it impact 90nm high-end release? That's the crux, it seems to me, the general feeling that if they go Ultra then they can't go 90nm high-end until 2nd quarter or so. . . But then I don't think we really know what their plans for 90nm high-end have been before now. All I have to go on is the CFO guy and his "90nm from 3rd quarter (2005)" (or somesuch) remark. . .

Edit: Otoh, maybe an Ultra can catch the 520XT that the ATI guys have been hinting at, at least initially, with det 80s. That used to be a (entirely legitimate) NV practice --mid-life driver juices on new architectures. Yeah, you can point at some corners cut pre-NV30 here and there, but mostly it was pretty good stuff.

 

[Tim Murray]

Has there been a respin or anything? 525 is a 20%+ bump in fillrate, which seems awfully high for your average Ultra model. Considering R520 doesn't have as much fillrate to begin with, that seems like a bit of a ridiculous jump.

 

[Geo]

525 was Jawed's number. I just said 480 was too low.

I haven't seen (doesn't mean there hasn't been --but I haven't seen it) any review that didn't get a GTX up to at least 490 in the oc section, on single slot cooling and spec juice. And several that went higher. Were those cherry-picked review models? Maybe, I haven't followed the forum reports at nvnews on what Joe User is typically getting.

 

[Rys]

Has there been a respin or anything? 525 is a 20%+ bump in fillrate, which seems awfully high for your average Ultra model. Considering R520 doesn't have as much fillrate to begin with, that seems like a bit of a ridiculous jump.

Some are close to 500 out of the box anyway, so it's not that big a leap if you ask me. And texel rate pretty much sucks as a measure of performance now. The question is what's 520's shader rate going to be like if you ask me

Along with that, the Quadro FX 4500 is the biggest hint you're getting to what's on the way. A dual-slot cooler already exists for that chip and G70 'GL' dun 'alf clock guv'

 

[trinibwoy]

525 was Jawed's number. I just said 480 was too low.

I haven't seen (doesn't mean there hasn't been --but I haven't seen it) any review that didn't get a GTX up to at least 490 in the oc section, on single slot cooling and spec juice. And several that went higher. Were those cherry-picked review models? Maybe, I haven't followed the forum reports at nvnews on what Joe User is typically getting.

Agreed, if Nvidia does release an overclocked-at-stock G7x variant, it would have to be somewhat faster than the current AIB offerings. 460-480 is way too low, I'm running 490 right now. So maybe 500 Nvidia recommended and ~ 530-560 AIB versions? I don't imagine it would take much extra voltage or cooling to take it from 490 to those speeds.

 

[Tim Murray]

I completely forgot about the higher-clocked AIB models, in which case yeah, 480 or so does make no sense. I'd guess it's going to be ~520 or so.

 

[trinibwoy]

I completely forgot about the higher-clocked AIB models, in which case yeah, 480 or so does make no sense. I'd guess it's going to be ~520 or so.

Mmmmmm, 520x24 vs 700x16 ?