The G92 Architecture Rumours & Speculation Thread
- Thread starter Arun
- Start date
- Status
That's interesting... 5000-6000 wafers is about one million chips before you take yields into account. I presume this is not NVIDIA's entire production, though.
PortalPlayer made its chips (including the GoForce 6100) at UMC. This makes me wonder whether their next-gen handheld will be at TSMC or at UMC; I had assumed the former, but now it looks like things may not be so simple...
This almost makes me wonder whether NVIDIA is going straight from TSMC 65nm to UMC 55nm... Or at least whether the two chips are slightly different in terms of specs or focus (notebooks?).
If NVIDIA wants to maximize their production capabilities for G92 so much though, it does make me wonder whether it will span a much larger price range than previously thought by some. 6C/384MiB/192-bit for ~$169 might get interesting... Interestingly, UMC's defects rates are higher AFAIK, which would turn out handy for such a part!
EDIT: On 65nm (200 chips/wafer and 75% yields including redundancy for arguement's sake), they'd need to sell each chip for an average of ~$60 to have 40% gross margins, assuming 300mm wafer costs around $5.5-6K. On 55nm, this would go down to $45-$50. FWIW, it is important to realize that 40% is their current GPU gross margin... Their average margins are higher because of the professional segment.
These calculations primarily serve to justify that a 6C/384MiB/192-bit SKU would make financial sense.
No, the card is well balanced. There is not one thing you can change that would yield an overall speed increase close to the increase in % you did to one setting. (typical settings and typical games)
No, the ONLY thing you can say based on this is that the card is not massively unbalanced. It does NOT prove that it's the optimal balance in terms of performance/dollar.
Mintmaster
Veteran
What's your point? Changing one thing by X percent never increases the total cost of the board by X percent either. In my post I was suggesting that 512MB with a 512-bit bus probably would make G80 notably faster (~15%), and would probably reduce cost.
If the increase was 1:1 with memory clock, then that means changing the core clock would have zero benefit, and they're wasting tons of silicon.
Mintmaster
Veteran
Not really. It just looks that way due to incomplete data and the stupid trendlines they drew.
I think you just answered your own question.Now if that's the case why the hell does the 4xAA graph flatten out at the exact same 700Mhz (at a lower performance level) if the workload should shift more onto the memory bus / shader core with AA enabled?
Remember that the GTX can do 4 AA samples per clock. Thus when you enable AA there are no additional chip cycles necessary and the extra load is entirely placed on the memory subsystem. There is also likely spare texturing ability w/o AF (depending on the shader), so again enabling it primarily increases the memory load as opposed to chip load.
R600's ROPs need 2 cycles to do 4xAA, so we have more chip cycles as well as bandwidth. There's also no hardware resolve, so this requires cycles that are "free" on G80. The same is true with AF, since R600 has far less filtering ability compared to G80. It didn't have idling filter units w/o AF, so the additional texture load now adds chip cycles.
The graphs show that AA/AF in fact don't shift workload onto the memory bus for R600. Instead, AA/AF adds notable workload to both the chip and the memory. This has been true on ATI products for quite a while. You'll notice that AA/AF hit on the Radeon 9700/9800 is almost identical to that of the 9500 Pro.
Blackiice2007
Newcomer
How much power will is consume the 8800GT ? i hope much less then 8800GTS .
True, you can´t say it is optimal balance as this would also depend on applications etc.
However 8800GTX shows that an increase of clockspeed usually gives more performance gain, then an increase in badwith (memory clock). It hsould be the other way round, if we want to believe that G80 as we know it would be much faster with just a hugee increase in bandwith.
No. The PCB would be much more complicated and more expensive (more layers as well probably), also not that nice for the memory layout.
vertex_shader
Banned
Sounds like original plan was a 6 cluster GPU before rv670 scared NV and enable one more cluster.
So NV not try to find more clock speed (as some rumor/user suggest), they just enable one more cluster and now the things looks like as usual.
Jawed
Legend
TSMC capacity constrained on 65nm? Didn't we get signs of this earlier in the year? Did AMD get the lion's share?
Jawed
Jawed
Legend
It could also point to yield issues.
But i'm betting on increased CELL and RSX orders as the main culprit.
Cell at TSMC?
http://www.fabtech.org/content/view/2469/
This was back in February. Who knows if they went ahead with it for 65nm ?
At least RSX is perfectly plausible, since it shares most of its "DNA" with TSMC-produced Nvidia G70/G71 GPU's.
- Status
Similar threads
