It appears that we can expect to see GM210 (aka Big Maxwell) next year on a 20nm fab process node with close to 10 billion transistors (!) and with exceptional performance, energy efficiency, and transistor density.
Look like Videocardz is definitively banned for get gpus themselves for reviews if they leak 2 reviews from other sites
Unlikely, since Big Maxwell is already on it's way, first sightings were barely 2 months behind GM204
Unlikely regarding what exactly? This calendar year ends in only ~ 3 months. And Big Maxwell will need to make use of the much more transistor dense 20nm fab process node to dramatically improve upon Big Kepler (there is no room to grow die size and transistor count much beyond that of GK110 on a 28nm fab process node).
Nvidia has been moving a sample of a GPU tagged as GM200 core (yes again). Earlier in July we noticed as well. This basically means that Nvidia has taped out the early A1 revision GPUs and engineering samples. This could mean a high-end launch Maxwell GPU as soon as Christmas IF wafer yields are decent enough.
...
This means that these processors will be HUGE in size, perhaps even over 620 mm²
http://www.guru3d.com/news-story/nvidia-gm200-(gtx-titan-ii)-spotted.html
However, that power draw apparently has gone down is quite an achievement - so "power/transistor" is only roughly half of what it was by the looks of it! Must have used a lot of low leakage, low power transistors, and I have no idea what other tricks (power gating, more fine grained clock gating...).
When was the last time that one company was truly ahead of the other in terms of architecture?
9800 (ATI), 8800GTX, RV770 and RV870, Maxwell ?
AMD better have an answer in the pipeline, otherwise winter is coming.
When was the last time that one company was truly ahead of the other in terms of architecture?
9800 (ATI), 8800GTX, RV770 and RV870, Maxwell ?
AMD better have an answer in the pipeline, otherwise winter is coming.
Wrong, they can/will grow die size.
That is true however Maxwell seems to look largely similar in overall structure so far to Kepler (at least on the hardware.fr diagrams...), I can't really see why the data would be a lot more local on average. But I guess improvements there are possible too even if you don't see them on the shiny block diagrams.Movement of data takes a lot of energy so optimizing data movement can save a lot of power.
Unlikely regarding what exactly? This calendar year ends in only ~ 3 months. And Big Maxwell will need to make use of the much more transistor dense 20nm fab process node to dramatically improve upon Big Kepler (there is no room to grow die size and transistor count much beyond that of GK110 on a 28nm fab process node).
Any reviews out yet? I thought the NDA ended 45 minutes ago?