Man from Atlantis
Veteran
i assume 13.3W@1600MHz(6400MHz effective), the bottom vrm readings belong to fbvdd, 1.6v is what 6Gbps memory chips use..
i assume 13.3W@1600MHz(6400MHz effective), the bottom vrm readings belong to fbvdd, 1.6v is what 6Gbps memory chips use..
huum what retain more my attention is in red :
maybe the card have OC and then go back to normal and so lowering consumption, i pain to see 75-100W difference can occurs in GT1 and GT2. Specially looking at the curve.. ( both test start high and then go down.. ) in GT3, you have directly a pic a peak down.
maybe the card have OC and then go back to normal and so lowering consumption, i pain to see 75-100W difference can occurs in GT1 and GT2. Specially looking at the curve.. ( both test start high and then go down.. ) in GT3, you have directly a pic a peak down.
It shouldn't cost much. What I like to know is how it's downclocked at 2D/Desktop. This card is using 6.0Gbps. The highest I've seen at default clock so far. I thought we needed newer vram ICs to do that (IE Samsung's new ones as point of reference)...
For 2D I'm not sure how that fits. But for TDP that's another story. It's still unclear how TDP is determined now that the card can auto overclock itself (game profiling??). The question would be if the 195W is based on just stock clock or does it include auto overclocks. However, the peeks and valleys do suggest that it's ramping up and down the clock rate. But I guess we will know more once the reviews are out.
Last edited by a moderator:
AMD's heuristics may be more conservative about ramping down, but we see that once a low-power period is long enough, it can drop below the 680.
The 680 looks like it has more granularity in terms of power states and it can switch faster.
The GPUs are physically subdivided differently, which may play into that.
The 680 has subdivided its compute areas into 8 sectors, and Tahiti may have not gone beyond two larger arrays.
The 680 may be able to ramp each section up and down more readily, while AMD's arrays may need to be more coordinated.
The 680 looks like it has more granularity in terms of power states and it can switch faster.
The GPUs are physically subdivided differently, which may play into that.
The 680 has subdivided its compute areas into 8 sectors, and Tahiti may have not gone beyond two larger arrays.
The 680 may be able to ramp each section up and down more readily, while AMD's arrays may need to be more coordinated.
This dont explain when in 3Dmark11 the gpu ask ( look fps ) is going harder in each GPU test, and the TDP decrease. if you look the AMD red line you see clearly the gpu TDP is increasing gradually from the start of the test to the end.. when the GTX is droping severely ( ofc if their curve is accurate ).
I can only think of some sort of game profiling for 3Dmark11. That may not be indicative to how the card will work in most other games. I guess we will have to wait and see.
Tnx! I figured it can't be more than around 15W.
TBH I think there's a lot simpler explanation - AMD drops to "low-power 3D clocks" or some such, as does 680, but 680 has simply lower clocks / voltages at that state
Man from Atlantis
Veteran
Then there's the new 3D Vision Surround, bolstered by a redesigned display logic, which addresses the two-display limitation of NVIDIA GPUs. You can now connect as many as four monitors to a GeForce Kepler GPU, enabling 3-monitor HD 3D Vision Surround setups. You no longer need more than one GeForce GPU to connect more than two monitors. The new 3D Vision Surround is said to work in conjunction with Adaptive V-Sync to ensure the center display has higher frame-rate (since it's at the focus of your central vision), at the expense of the frame-rates of the two side displays (since they're mostly at your peripheral vision). This ensures there's a balanced, high-performance experience with multi-monitor gaming setups.
TPU
At 1st guess it looks like throttling. But we need to see charts on frame rates, time to render, etc to see whats really going on.
Anything is possible until the official benchmark release date.
A1xLLcqAgt0qc2RyMz0y
Veteran
Which is negated when loaded.Pic regarding the link posted earlier regarding the leaked benchmark results...
I'm not sure how accurate this is but what got my attention was the consumptions rates when their is no load.
AMD will have lower consumption when idle but is saving 20 watts when idle and burning 50 watts more when active really a plus.
We are of different opinions then :smile:. I don't know of many who game that often in a 24 hour period where the difference in load would matter that much. However, in 2D mode when some can go days without playing a game is a different situation.
I gave a possible explanation.Well, since the actual layout of the RF apparently doesn't match the supposed number of SIMDs, it may very well be a plausible assumption. And more over, if Kepler treats the RF as a single address space, like in Fermi...
4 Schedulers, each having access to a set of 3 vALUs (16 lanes, instruction issue blocks it for the next cycle), 16 L/S (issue blocks it for the next cycle) and 8 SFUs (issue blocks it for the next 3 cycles, i.e. issue every 4 clocks max). One SMX would be an aggregation of four GF104 style SMs without hotclock (hence halving the scheduling needs) with shared TMUs and local memory. An evolution of the G80/GT200 TPC layout with all the Fermi sauce.
To get maximum utilization, the processing of the pixels in a quad has to get out of sync. I really doubt this happening with dynamic scheduling right now.
When trying it statically, it has basically the disadvantages of a combined SIMD8-VLIW8 approach (you described it more as a dual issue VLIW4, but it is not going to work this way) and would try to conceal this by processing of a quad and not a pixel in each (VLIW)-vector lane. For some problems it might work okay, but generally it is not really compatible. Just imagine the hassle when a branch diverges within a quad. How do you compile for that?
And with dynamic scheduling (without VLIW) you would actually need an 8 way scheduler for it.You would need to split each 32 element warp in 4 quarter warps (each quarter warp has one pixel from one of the 8 quads in the warp) and schedule the quarter warps individually. And you don't get any performance improvement compared to my suggestion above, but you need far more complex scheduling hardware.
Man from Atlantis
Veteran
I never indicated any exact wattage numbers. However, it is what it is based on that chart if it's true.
Are you talking about this:
Man from Atlantis
Veteran
it's not indicating the idle power usage, it indicates previous power state than 3D clocks.. so you cant conclude a result from it.. GF114 is bigger and manufactured on a worse process, now there is GK104 is smaller and manufactured better process tech.. Tahiti is much bigger, yet draws less than Barts.. i dont expect anything above 15W idle.. and anything less than 15W shouldnt be a problem for money or heat..
chart shows 200W idle according to your say, seriously 200W idle ??
