itsmydamnation
Veteran
i have lurked @ B3d since about then 2nd or 3rd interation of the R600 prediction thread
, i cant agree with your self assassment.note im australian which means i go for the underdog which is generally AMD
NVIDIA Fermi: Architecture discussion
- Thread starter Rys
- Start date
i have lurked @ B3d since about then 2nd or 3rd interation of the R600 prediction thread
, i cant agree with your self assassment.note im australian which means i go for the underdog which is generally AMD
With perf/watt being so much better with GPUs in general than with CPUs, i think any vendor who can provide an x-factor against CPUs doesn't necessarily need to care if that x equals 12 or 11 or 10 or even 9. And with each card possibly sporting a six- and an eight-pin-connector next to the 75 watts from the PEG slot, staying inside the power budget of 225 watts would be nice but not a necessity IMHLO.
Sure, given the prices asked one would expect that you'd get the cherry-picked chips with the higher price tags - but i guess a large portion of that higher price tag and the customers willingnes to pay the bills comes from a) the software ecosystem support (which is in case of commodity graphics cards only or mostly drivers) plus the real money, professional users can save by changing their hardware whereas gamers only get more Fps (or higher image quality).
And surely Nvidia will go out of its way to not sell "imperfect" GPUs, but "GPUs that fit a given specs".
Do we really know this for sure yet wrt to gaming performance?
Which is what GTX260 was to GTX 280. And if you were around at the time of 4870's launch on B3D forums, 260 was said to be DOA on it's then launch price. And how well did it turn out for nv?
Again, I'm not assuming that what I called "GeForce 360" (384 SPs, etc, etc) is just a full Fermi chip, with some units disabled. I'm guessing that that part (with units disabled) is actually a GeForce 370, with 480 SPs, one or two ROPs disabled, etc.
I hope that this time, NVIDIA doesn't go around with "Core X" and/or "Reloaded" BS suffixes, to distinguish models from one another.
i have lurked @ B3d since about then 2nd or 3rd interation of the R600 prediction thread
note im australian which means i go for the underdog which is generally AMD
Actually the underdog is nvidia
AMD is only an underdog to Intel.
Via is a real underdog you ought to rush out and embrace them.
Actually the underdog is nvidia
An underdog doesn't have more market share than its competitor
Right, now that you're mentioning it…Of course not. This is a speculation thread, remember ?
Seriously: We shouldn't forget the possibility that unit for unit and clock for clock, the nvs new arch could, theoretically, be slower than before. It's not even without precedence: Geforce FX.
Right, now that you're mentioning it…
Seriously: We shouldn't forget the possibility that unit for unit and clock for clock, the nvs new arch could, theoretically, be slower than before. It's not even without precedence: Geforce FX.
Sure, but why is there a need to always assume the worst for NVIDIA ? This phenomenon, which is usually what Charlie does, isn't really logical.
Unless you want them to fail, one expects that a new architecture brings actual improvements over the past one. GeForce FX was bad yes, just like R600 for ATI, but those failures can't be the starting point for the speculation of a new part by any of these companies.
If the fermi (or it's salvage part) is neck and neck with 5870, then nv's margins in the consumer gaming high end market are toast.
Agreed, but again, that's not what I'm saying.
This is why (look below).
Funny you're assuming that nVidia's ALUs will scale much, much, much better than ATI's (in your case ideal).
That's like counting architectural improvements twice when you don't even know how Fermi's modifications will impact how it takes on the graphics pipeline at large. Oh well, people are already saying Amen so others assume it's true.
At 4200Mhz (same as Rys' speculation) the 320-bit GTX360 will have 6% more bandwidth than the 285.
And again, if this is a salvage @575Mhz core (reasonable spec wrt GTX260a) or so, you have again a whooping 6% more texel and pixel fillrate.
The only point here is the substantial shader power increase- it's 1.6x base ALUs and say, 1600Mhz (VERY optimistic imo for a firstrun salvage part, I'd put my money on 1400Mhz instead) - 8.5% more shader clocks vs GTX 285.
73% more shader power but nearly flat fills beside Z. My money's on the 5870 being 10+% faster at least (and this is the part of me that's not bullish on future Catalyst releases- the other 99% of me is)
I don't think so.
The scaling in NVidia has been even lower than in ATI parts. e.g.: G94 is almost an half G92, but it performs much higher than half G92. That is due to:
- non linear ALU nr. - to - performance scaling
- other portions of the gpu not increased in the same proportion (i.e. bandwidth in the G94-G92 case)
Not sure why. The scaling of two very similiar AMD parts has little bearing on the scaling of a new Nvidia architecture vs the old.
Do you have numbers to back that up? I always see people talking about margins but the actual financial results of the companies should trump assumptions and speculation no? AMD's die size advantage looks to be smaller this round so it's kinda hard to tell who will have better margins until we see Fermi performance and pricing. Last I checked AMD was selling 2 RV870's and 2G of GDDR5 for $600.
Yeah, Fudzilla itself doesn't count as a source and the story doesn't source anything!
Remember, Fudzilla just told us yesterday that Fermi was coming in January....if they're right on that one mebbe I'll give them a bit more credit on this.
leoneazzurro
Regular
And, in the most of the cases, even the GTS250 has not half the FPS of the GTX285, except in corner cases where only bandwidth matters.
I think that's just based on the assumption that if you have a die size which is still ~50% larger than what your competitor has, and that the chip manufacturing costs are significant, this should probably make some difference in your margins, if you're forced to sell at a similar price than the competition. Though maybe this time PCB doesn't have to be more complex (the part with disabled units could have only 256bit memory bus, even with 320bit it might not be much more complex though it would need more memory chips). Other than the chip and the memory interface width, there just doesn't seem to be much reason why one or the other card could be cheaper to build.
Though I guess if the top part is fast enough so the lower part can have quite a few units disabled and still slightly beat the HD5870, this would allow nvidia to charge quite a bit more for the top part improving margins significantly without making it a niche product.
GTX 285 vs GTS 250
Fillrate: +76%
Texturing: +10%
Bandwidth: +126%
Flops: +51%
Performance advantage ranges from 50-80% depending on settings. How you can paint that as worse scaling than 4870->5870 is beyond me.
Jawed
Legend
This gap only gets bigger as the performance of the enthusiast class card increases. It doesn't matter who makes the card
Things like amount of memory, cache size/behaviour, driver efficiency are in NVidia's control, so there are opportunities to minimise the problem. e.g. if setup rate is a serious bottleneck in HD5870 but NVidia tackles this head-on...
20% is a serious amount. It cautions that we should be thinking in terms of a similar shortfall needing to be applied to GF100. We can only wait to see what it turns out to be. It'd be silly to assume it's 0. I agree that neither GTX285 nor HD5870 are useful baselines. But if we cautiously assume 80% of what theoreticals indicate based on estimated unit counts and clocks, we'll be better armed to understand how it does scale when it appears.
I don't know of a basis to estimate this gap. HD5870 might be a refresh too far and a completely revised architecture fixes it, e.g. brings the gap down to 5%.
Separately, D3D11 games might show a marked reduction in the size of this gap, simply because with things like TS and CS, considerably more-efficient use of memory is possible.
Jawed
AMD will still have a ~50% die advantage. Not to mention that if 5850 can turn a profit at $259 (full supply price), so can 5870 in worst case (if 5850 is eol'ed), but not nv.
You may say that about the HD 5870 (assuming worst case for NVIDIA), but definitely not about the HD 5850. It's a salvage part of the full RV870 chip and NVIDIA certainly won't be competing with it with a salvage part of the full GF100 chip. I'm guessing a new chip for a "GeForce 350" or "GeForce 340", with 256 SPs, half the ROPs and half the TMUs on a 256 bit is probably what will.
Maybe that's the second chip that will be released (and not the GeForce 360), according to Fudzilla. This means that the release would have the high-end and the mid-end portions of the market covered.
Going out on a limb and using G94 as an example, which is roughly half of G92 in every way (just like I'm speculating this "GeForce 350" to be, when compared to the full GF100 chip):
G92 @ 65 nm = 324mm2
G94 @ 65 nm = 240mm2
Assuming GF100 @ 40 nm = 480mm2 (as seems to be the most common speculation)
The chip that powers the "GeForce 350" (GF104?) @ 40 nm should be around ~355mm2.
If it competes or even beats the HD 5850 overall, then this segment will surely be interesting to follow.
