Nvidia G71 - rumours, questions and whatnot

Excuse my ignorance, but is it possible that nV has further beefed up the secondary ALUs? Would this be effective (isn't the secondary ALU tied up primarily with tex ops?)?

 

Tied up yes; primarily no.

More ALUs per SIMD channel wouldn't had been an option IMO. For the future de-coupling ALU from texture OPs makes way more sense.

 

I make it 24x625MHz PS and 10x580MHz VS.

Jawed

 

While it would be an option, I don't see any reason for different clock domains anymore.

 

Yeah, I was about to link that. 24 pipes seems so logical from a respin and margins POV, but I'm still wary given the intense hype for 32. Plus, 655MHz doesn't square with any of the given #s. Is Sven trying to out-Fuad Fuad? Maybe he deliberately shortened the fillrate and bandwidth figures. 655*24 = 15.72B, so pretty close.

Not sure how to take the vertex #. Using the same formula as NV does on the GTX 512 page, it's either 8 VS at 725MHz or 10VS at 580MHz. I'm favoring the former, but then we've heard 10VS bandied about as often as 32PS, and I believe the GTX 256 was the only G70 to run a higher VS domain clock....

The question still remains if they've bumped the ROPs to 24 or stayed at 16. I'm not sure if more ROPs are due given slower stock RAM clocks.

 

If these rumoured specs are accurate, G71 = G70 with 100MHz boost on the core and slightly lower memory bandwidth?

So other than a modest performance boost over G70 in some titles, ultimately the main attraction will be a lower price tag

 

I dunno, I won't complain about a faster, cheaper GTX. Given how the GTX-512 compares to the 1900, NV'll be in the hunt, and their CFO probably wouldn't object to a tiny die.

Would extra ROPs help AA performance, even theoretically?

 

Is that 7600GT or GS?

 

Parhelia's memory interface had 4 channels. So 4x64.

 

Should we rename it the M(argins)71 for them?

A few factors to consider, if taking the numbers we've been hearing about today seriously:

1). Jen-Hsun, not satisfied with his record 40% margins, has been predicting a rise to 45% margins. While they have been quick to insist that improving margins cannot be accredited to any one factor, having their flagship go from ~335mm2 to ~225mm2 certainly couldn't hurt!

2). Re 7600, going from NV43 to 90nm 8/12 won't hurt either.

3). The whopping 45% core clock jump from 7900GT @450mhz to 7900GTX512 @650mhz for same# of quads suggests strongly to me very aggressive clock-binning at the top. Is there any comparable delta in core clock rate in recent history between the top end sku and "one down"? None are coming to my mind.

4). Looking at #3 above, and taking 16/24 seriously, one has to look at 7800GTX512, and NV's statements in the first 1/2 of 2005 re everything being 90nm in the second 1/2 of 2005 (let's assume FY, so from Aug 1), and take very seriously the possibility that we are looking at a part that is signifcantly later than planned and 7800GTX512 was a stop-gap while they kept trying to work the clocks higher on G71.

5). Looking at current X1900xtx vs 7800GTX512 benchies across a wide range of games, aside from B3D'ers natural interest and joy in new stuff/features/changes etc, the fact of the matter is that a die shrunk G70/71 @ 650mhz might be "boring" to us, but it will be a darn competitive part for the next few months at least. And after that, Vista-obsession will take over and if any new games appear on the scene to embarrass them in Dynamic Branching, Alu/Tex ratio, or what have you they will be close enuf to NV50/G80 to say "All the benchmarks show our current part is very competitive, etc. . .and for 'next gen' games, well stay tuned for our 'next gen' part, coming up shortly."

6). And given #5 above, and a obvious disinclination to make major changes this time, why muck about with another two quads? It doesn't give you HDR+AA. It doesn't give you HQ AF. It makes you a few fps faster in benchies you're going to win anyway and a few fps faster in benchies you're going to lose anyway. So why bother to add cost for no real benefit?

All of which seems to suggest I'm taking 16/24 very seriously, and indeed I am, considering that a month ago I was just this side of certain it would be 24/32.

But I'm not making a call on this. Instead, I'm back on the fence. Why? Because there's been just a little too much tomfoolery for my comfort, as exemplified by that 7600 package in a 7900 pcb at dailytech. Frankly, I can't think of any reasonable reason for NV to be mucking about with an elaborate misinformation campaign at this point (for instance, say, imminent R590 launch to trump them would be an example of a reasonable reason), but there's something funky in the air and I'm not going out on any limbs with that scent in my nostrils.

7600GT, on the other hand, I'm feeling comfortable is 8/12. We'll see. I may be wrong, but I'm not unsure.

 

Or conversely, an attempt to guarantee massive availability at the bottom?

 

45% percent is a lot, but not uncommong to nV, right?
430 to 550 is allready close to 30% and seen the massive improvement the 512 is over the GTX, a GTX at 650Mhz would be powerful, very powerfull.

 

430 to 550, right, but also 600 to 850 memory speed. there will be no big increase of memory speed. So I wouldn't expect that much more performance.

 

Let's wait for the specs first

IMHO the high-end part will be about just as fast as X1900 +/-5% in most games and priced a little bit below initially.

EDIT: and in games which are not that limited by shader power or dynamic branching, it will be ~30-40% faster than X1900.

 

I find it weird that they would take a step back on the memory front (850 for thr 78-512 back to 800 for the 79-512.)

At least the acces times should go down right? Latency could be a deciding factor

 

The gap between 7900 GTX and GT in MHz for the core just seems too large to me. I can't equate the speed of the GTX at 655 with the values given either. I think I shall just sit and wait, this speculation is not getting much more definite at this point I think.

As suggested above as always it is probably going to be " there or there abouts " .. which is not very exciting to be honest, but probably accurate !

 

EDIT: Missed the post which already had the dailytech link.

 

The 7800GTX 512 shortage was affected by shortage for the ultra fast memory, so I guess they want to avoid that.

 

I always considered the GTX512 to have more bandwidth than it actually needed (in a relative sense).

7800GTX 256 = 10.3 GTexels/s with 38.4 GB/s
7800GTX 512 = 13.2 GTexels/s with 54.4 GB/s

A ~28% increase in fillrate and a ~42% increase in bandwidth.

Now would I add theoretically 14% more fillrate to the GTX512, I'd reach 15.05 GTexels/s.

So far rumoured specs for the 7900GTX suggest 15.6 GTexels/s with 51.2 GB/s bandwidth.

 

No there aren't. Dailytech reports 900MVertices/s for the GT which means 8*450 (no different clock domains). It sounds like being only within a breath's difference from the 7800GTX 256.

Before I know the price though for that one and it's relativity to the price of the 7900GTX, it's too early for conclusions. It's aiming to replace the 7800GT in a given timeframe and those are currently available starting from $285.