G80 rumours

That's a bus-width hint or a framebuffer size hint?

Edit: Okay, the actual phrase that popped into my brain was "384-bit?"

 

Would it implicate it will have 1 gig vidmem on board ?

 

1024/12 = 85,333MB per chip...

768MB (or 384bit mem-bus) sounds more reasonably, but still curiously.

 

Btw, here's what sireric said about 384-bit generically, tho obviously there is no guarantee that NV engineers would necessarily be thinking about it the same way: http://www.beyond3d.com/forum/showpost.php?p=692169&postcount=53

I've been curious for some time about whether 384-bit would be a feasible/interesting thing --in fact since XGI did that 192-bit bus.

 

Sounds interesting too, but I asummed this next gen would have 1 gig, as T Sweeny once said to fully experience UT2007 in it's full glorie, 1 gig vidmem would be requiered.

 

Well, it seems doable all right, but with the extra memory bandwith that GDDR4 provides, I wonder whether nVidia will go to the extreme on this account already.

 

We might be talking about 256-bit bus and an additional dedicated 128-bit bus or 2x64-bit busses according to that post from what I get. Hmm, sounds quite interesting...

 

I get the distinct impression that because of design timescales, G80 will use GDDR3 (at least first of all). Therefore, if you assume the chip to have a goodly bump in theoretical render power and rates, to feed that with GDDR3 (which stops shy of 1000MHz remember) would reasonably require a larger bus width than we see today.

So I'd say GDDR3 with a > 256-bit bus is certainly on the cards. I'd been working on 512-bit myself (> 100GB/sec/chip!), while doing the usual chip workup stuff before all the juicy spec tidbits arrive, pre-launch, but 384-bit gets NVIDIA into the ~86GB/s/chip realm, which is close enough right?

 

A distinct impression? Why?

I will gladly accept that. GDDR3 also have lower latency BTW.

 

Samsung simply aren't providing anyone other than ATI atm, from what I can tell, and haven't been in MP for that long either (less than 2 months now). NVIDIA will require a good amount of physical DRAMs for bring-up attempts, never mind 'decent' sampling, and remember the chip hit silicon when GDDR4 was simply impossible to get.

At least current work with G80 is with GDDR3 (IMO, but an O that I'd put money on).

 

Good points. What I want right now is a G80 paper launch. Enough of this guessing, dammit!

 

Well, from what everything nVidia has been saying over the past few years, it seems that they have gotten enough positive feedback from their hard launches that they would like to continue with the strategy.

 

Maybe it's a 256-bit bus to conventional local RAM plus a 128-bit bus to a pool of memory dedicated to:

• constant buffers
• post geometry shader cache
• stream-out

Jawed

 

I don't see any reason to split up the memory bus. While the costs of high-end graphics cards today are certainly high enough to support a wider memory bus, it makes much more sense to extend the current crossbar memory controllers than to have a dedicated bus. Dedicated buses tend to suffer from inadequate use.

 

Maybe a 128-bit bus to something like EDRAM?

 

1. I would expect any eDRAM bus to be much wider, otherwise what would be the point?
2. We are talking about a possibly strange number of external memory chips here, after all

 

External as with Xenos?

Hey, just keeping the mill going...

 

geo rummages around in B3D Closet of Stuff. . . "Now where did Wavey leave that #$%^ frying pan?!"

 

Given X1950 is 64GB/s, yeah, 86GB/s sounds like a nice bump. It does make me wonder a few things tho, like will they consider bumping the ROPs now with the new channels? We know that ATI had some issues optimizing memory access for the original 3-quad X800Pros, for instance, if I recall correctly. Tho NV has had the decoupled ROPs for some time now, and that's possibly helped them a bit here?

And what are the implications for a hypothetical G8600? Seem likely it'd stay at 128-bit? I'd think not. But will it be 256-bit? Or 192-bit, to leverage anything they might do to opt for 384-bit in its big brother?

Also does 12 chips have implications for total framebuffer size? Can they still get to 512MB with that, or will we be looking at an odd size (for all us power of 2 bigots) fb too?

Altho all of us need to recognize/acknowledge we've gone a long ways from one rumored sighting of a 12-chip board. That's fine, because that's what we do around here, but until I see a piccie myself, I'm not headed to the bank. . .

 

Btw, this thread is 28 pages now. . . I'm wondering if it's time to pinch it off and start a new one. . .