It looks like the first Maxwell GPUs (GM107-300 & GM107-400) are being fabricated with a 28nm process, as found in preceding Kepler chips. We recently reported that the 750 Ti is rumored for a February 18th release date, and noted that the 750 Ti would be nVidia's debut of the Maxwell GPU line.
Source: VideoCardz.
Anyone who pays attention to the industry knows that nVidia tends to lead with their flagship GPU, releasing lower-SKU cards much later in the year. This time, though, it seems the 750 Ti is opening; we theorized that this is because the 750 Ti is more accessible to the greater whole of PC gamers, and thus has more potential to gain recognition. With this new information, though, I think the reasoning is different: I think nVidia is waiting for TSMC's fabrication capabilities to ramp into 20nm process before shipping higher-SKU devices.
From what we know so far, the new GPUs are named the GM107-300 and GM107-400. The GM107 GPU is a fusion of the GK107's 128-bit memory bus and the GK106's CUDA core count. The GM107-300 will ship with 768 CUDA cores, with its bigger GM107-400 shipping with a targeted 960 cores. It is reasonable to assume that the GM107-400 will be used in the GTX 750 Ti video cards while the GM107-300 is deployed in lower-power cards, like the GTX 750 (and discrete GPUs that don't require dedicated power lines).
NVIDIA GM107-300 & GM107-400 Specs (GTX 750 Ti Specs)
| GPU | GM107-400 | GM107-300 | GK106-400 | GK107-450-A2 |
| Fab Process | 28nm | 28nm | 28nm | 28nm |
| Die Size | 156mm^2 | 156mm^2 | 221mm^2 | 118mm^2 |
| Memory Bus | 128-bit | 128-bit | 192-bit | 128-bit |
| CUDA Cores | 960 | 768 | 960 | 384 |
| TMUs | 80 | 64 | 80 | 32 |
| ROPs | 16 | 16 | 24 | 16 |
| TDP | ~75W | TBD | 140W | 65W |
| Likely Found In | GTX 750 Ti? | GTX 750? | GTX 660, 660 Ti | GTX 650 |
| Release Date | February, 2014 | February, 2014 | September, 2012 | September, 2012 |
It would appear that both existing GM10X GPU iterations are using the current 28nm fabrication process on a 156mm^2 physical die size, which is shrunken nearly 30% from the GK106 (die space has been reduced). I think this is the real reason we're seeing lower-end GPUs leading the higher-end stuff this time: TSMC is still pushing out 28nm process silicon, so we won't see a bombastic "GTX 800" Maxwell release -- 880 equivalent and all -- until 20nm process is available. That's my guess. This will be a much softer launch to drum up some interest and get discussions on the new architecture going, but won't break the doors down as we'd expect from a high-end flagship card.
And that initial discussion is important: Maxwell uses less die space, has a lower (or equal, in some cases) TDP, and has significantly bolstered processing power in the form of additional programmable CUDA cores.
Other sites have called the GM107 "another Kepler refresh."
NVidia's GTC (Graphics Technology Conference) is in late March; it is highly likely that we'll be able to learn a lot more about the actual underlying architecture of future GPUs at that time, though I would anticipate a GPU launch before the conference begins.
- Steve "Lelldorianx" Burke.