GP106 speculations

[Mopetar]

If you're talking about bus power cards, that's limited to GP107.

We've seen the 1080 -> 1070 transition, 180W -> 150W. Another cut will result in around 125W or so, well above "bus only power".

That's why I said if it happened it would be a harvested part. Turn off 25% of the SMs and drop the clocks a little and there's no reason that GP106 couldn't run off of board power.

If they take long enough to launch it, they could even use GDDR5X memory for power savings. I'm sure there's some modules to be harvested there as well.

 

[jpiniero]

We've seen the 1080 -> 1070 transition, 180W -> 150W. Another cut will result in around 125W or so, well above "bus only power".

nVidia did release an GM204 "960" for OEMs only (1280 cores, 924 Mhz) that was only bus power, so it's possible that the second cut GP104 is a similar idea.

 

[JDG1980]

That's why I said if it happened it would be a harvested part. Turn off 25% of the SMs and drop the clocks a little and there's no reason that GP106 couldn't run off of board power.

Sure, why not? They did with GM206 (GTX 950). That too was a 25% cut.

If they take long enough to launch it, they could even use GDDR5X memory for power savings. I'm sure there's some modules to be harvested there as well.

IMO, GP106 should really have a 128-bit GDDR5X bus. In the long run it would be cheaper for Nvidia and AIB partners than a 192-bit GDDR5 arrangement, even if it costs a bit more at this moment. It would also save on TDP. The cut SKU could have standard GDDR5 (it would be half a 1070, the same way that the full GP106 would basically be half a 1080).

 

[iiiankiii]

I really like the GTX 750TI in terms of power draw. Having it able to fit in a lot of OEM PCs made it a pretty good card. However, the performance was a bit lacking. How powerful do you think a GTX 750TI replacement will be?

 

[tviceman]

GP106 has a 256bit interface.

No it absolutely will not be. GM206 was 128-bit, GK106 was 192-bit, and GF116 was also 1920-bit. There is absolutely no reason for GP106 to have the same bus and potential bandwidth as GP104.

If GP106 has 1/2 the cores as GP104, then it will also have 1/2 the bus width, just like GM206. If it's not exactly half and the cores are grouped differently, like GK106 was to GK104, or GM107 was the rest of Maxwell, then I suspect it will be 1536 cores (60% of GP104) with a 192-bit bus. The rumored die size is between 170-200mm2. At 170mm2, that would support the 1/2 everything GP104. At 200mm2, that would support the 60% cores and 192-bit bus of GP104.

 

[tviceman]

Sure, why not? They did with GM206 (GTX 950). That too was a 25% cut.



IMO, GP106 should really have a 128-bit GDDR5X bus. In the long run it would be cheaper for Nvidia and AIB partners than a 192-bit GDDR5 arrangement, even if it costs a bit more at this moment. It would also save on TDP. The cut SKU could have standard GDDR5 (it would be half a 1070, the same way that the full GP106 would basically be half a 1080).

TL;DR There is no history or precedent Nvidia has set for it's x06 sku, as it's been different every time since Fermi.

It depends on so much more than what we know. You could be right about the pricing short term / long term of G5X on 128-bit vs. G5 on 192-bit, but we have no idea of G5X's availability / production to support multiple products, one of which being a mass-market ~$200 card. We also have a history of X06 cards not following the same pattern set forth by Maxwell (i.e. 50% cores and bus of the x04 sku). GK106 was not 50% of GK104, and GF116 was not 50% of GF114. Also, if Nvidia does follows the Maxwell that pattern of cutting the next chip in half of it's bigger sibling, that would eventually make GP107 barely faster than GTX 950, only representing about a 33% increase in speed over GM107.

 

[R0H1T]

I really like the GTX 750TI in terms of power draw. Having it able to fit in a lot of OEM PCs made it a pretty good card. However, the performance was a bit lacking. How powerful do you think a GTX 750TI replacement will be?

At least 2~3x as much as the 750Ti, with the custom cards being closer to 3x in terms of performance, while the passive cooled versions should still be ~2x as fast.
This is wrt both the camps as AMD also seems to be dipping their toes in the 750Ti territory with a P11

 

[iiiankiii]

At least 2~3x as much as the 750Ti, with the custom cards being closer to 3x in terms of performance, while the passive cooled versions should still be ~2x as fast.
This is wrt both the camps as AMD also seems to be dipping their toes in the 750Ti territory with a P11

No way, man. A GTX750TI is actually a pretty fast card. 3X faster ~ GTX980. I highly doubt that. I can see the GTX750TI replacement being around GTX770/R9280X terrority.

 

[R0H1T]

No way, man. A GTX750TI is actually a pretty fast card. 3X faster ~ GTX980. I highly doubt that. I can see the GTX750TI replacement being around GTX770/R9280X terrority.

There are GTX 950 without power connectors, the 950 is also substantially faster than a 750Ti.

So, I do expect the next gen of cards to deliver ~2x performance over the 950, without a power connector.
Therefore 2~3x 750Ti is very much doable especially in DX12, even at a TDP of 60W, however do note that I never said perf/w & just absolute performance :sneaky:

P>S> the 3x faster than a 750Ti is taking the lesser card as base, just as the 950 is ~43% faster than a 750Ti whilst the latter is 30% slower at 4K, using the TPU charts above. The 980 is an insane OCer but more importantly at higher resolutions it's around 3.5x as fast as the 750Ti, also the next gen cards will do much better in DX12 & that's where I see ~3x being made possible.