NVIDIA Maxwell Speculation Thread

Even if denver is unified, it may not be used in current driver.
However, I think that denver will not be included in early maxwell because there was discrimination of SM35,SM30 in kepler.
But I think if extreme story nvidia puts TegraK1 for five 64bit together even if GPU similar to GM107 is completed.

(5*k1(5W)=25W) wow

I look forward to denver.
 
The most impressive part about GM107 is its perf/W ratio. I've no clue how any Kepler could reach such a ratio. As a reminder:

GTX650 (GK107) = 12+ GFLOPs FP32 /Watt
GTX680 (GK104) = almost 16 GFLOPs FP32 /Watt
GTX780Ti (GK110b) = 20+ GFLOPs FP32 / Watt

It shouldn't be much of a surprise that that ratio on GM1xx would be at least twice compared to any GK1xx predecessor; and yes of course FLOPs aren't everything.
 
I don't know if games will scale that well though, the bandwidth is just so low. Early rumors said an integrated Denver core is somehow helping to alleviate bandwidth constraints, but the past several pages of comments seem to doubt a CPU core is integrated, let alone able to function in such a manner.

Sorry but how an integrated core may alleviate bandwidth constraints?
 
The most impressive part about GM107 is its perf/W ratio. I've no clue how any Kepler could reach such a ratio. As a reminder:

GTX650 (GK107) = 12+ GFLOPs FP32 /Watt
GTX680 (GK104) = almost 16 GFLOPs FP32 /Watt
GTX780Ti (GK110b) = 20+ GFLOPs FP32 / Watt

It shouldn't be much of a surprise that that ratio on GM1xx would be at least twice compared to any GK1xx predecessor; and yes of course FLOPs aren't everything.

All I'm saying is that binning is an important variable in power-efficiency, and in this case it's an unknown variable. Of course it's entirely possible that GTX 750s are binned just the same as desktop GK107s, which would make them quite impressive.

Is the process situation clear? I mean, could it be that Kepler was made on 28nm HP and that Maxwell is on HPM or something like that? To be honest I didn't expect this much of an improvement on 28nm.
 
Probably maxwell may reduce useless data transmission and calculation by making out the schedule smart in cpu.
This becomes the energy saving.

ex

dynamic warp formation
dynamic warp subdivision
 
All I'm saying is that binning is an important variable in power-efficiency, and in this case it's an unknown variable. Of course it's entirely possible that GTX 750s are binned just the same as desktop GK107s, which would make them quite impressive.

Is the process situation clear? I mean, could it be that Kepler was made on 28nm HP and that Maxwell is on HPM or something like that? To be honest I didn't expect this much of an improvement on 28nm.

I wouldn't exclude the possibility of HPm, however how high would you rate the difference to 28HP? More than 15%?

Oversimplified assume you gain from generation to generation 80% in efficiency. Let's say you get 30% out of that from the process (best case scenario) the rest comes from where exactly?
 
Becoming it is an architecture level, but electric power saving is possible.
 
Last edited by a moderator:
Becoming it is an architecture level, but electric power saving is possible.

Not to as high degrees; if yes then we've been facing sluggish engineering for decades for about everything out there.
 
As for you, me sorry not understand becoming it too.

I'd expect the next in line chip around the start of this summer and the successor to that around fall. The question remains from which point it becomes a cost necessity to move to 20SoC. Considering GM107 has roughly 32% more die area than GK107, I'd dare to speculate that they might even keep "GK104" or else the performance part on 28nm for its first appearance.

http://semiaccurate.com/forums/showpost.php?p=207239&postcount=249

I am sorry for quoting you, Ailuros, from another forum (Semiaccurate's), but I could not be bothered with registering there, and I have a question:

If GM107 is using 28HPM, would it be viable for a chip sized say 300mm2 to use it as well?
 
http://semiaccurate.com/forums/showpost.php?p=207239&postcount=249

I am sorry for quoting you, Ailuros, from another forum (Semiaccurate's), but I could not be bothered with registering there, and I have a question:

If GM107 is using 28HPM, would it be viable for a chip sized say 300mm2 to use it as well?

I'm not even sure it's using 28HPm; I can't obviously exclude the possibility at this stage.

As for 300 or more mm2 I haven't read anywhere that there's such a threshold for 28HPM but I'd love to stand corrected.
 
I wouldn't exclude the possibility of HPm, however how high would you rate the difference to 28HP? More than 15%?

Oversimplified assume you gain from generation to generation 80% in efficiency. Let's say you get 30% out of that from the process (best case scenario) the rest comes from where exactly?

Sounds about right for HPM: http://images.anandtech.com/doci/7082/Performance400.png But that's the only concrete data I'm aware of.

As for micro-architectural gains, I didn't expect Kepler to leave this much room for improvement. Previous iterations were different and Fermi, for example, was a clear case of low-hanging fruit with its hot clocks. Whatever NVIDIA did to make Maxwell more efficient, I hope they go into some detail to explain it, as I'm very curious.
 
Why shouldn't there be any low hanging fruit for a Kepler or any other future architecture? Have we reached perfection already & any future research is in vain? It'll be almost 3 years from the Kepler introduction until all Maxwell cores unroll. If it wouldn't be for process technology to slow things as much down, Lord knows where we could have been today.

The signs have always been there. Their roadmap concentrates on Tesla DP FLOPs/W, but Kepler is marked for roughly 6 GFLOPs DP/W and Maxwell at at least twice as much always in a rough ~225W TDP envelope. Assume Maxwell top dog is also 1:3 DP-SP you're looking at 9-10 TFLOPs FP32 in 250W TDP.

Isn't Intel also promising over 3 TFLOPs for KNL? Where's the surprise exactly?
 
gtx 750 in preorder in Italy (arrives 7 february) at €150 vat included (22%)

6ZtJ3TB.jpg
 
Back
Top