Indeed, looks fake to me. Why on earth would they give their top card lower bandwidth than anything high end from the previous gen? If that card does exist, its more than likely a 770 replacement, not a 780 ti/titan black replacement.
Nvidia have had trouble since the 480gtx getting a large die size part on a new process remotely quickly, hence the 680gtx 300mm^2 part instead.
It's also listed as a 7.9 billion transistor part while a 780ti is a 7.1billion transistor part.
New process = roughly double transistor density, high end to high end on different processes usually around double the transistor count.
A new process with 1.9x the transistor density yet barely 10% more transistors screams undoubtedly, midrange.
If you have say low/mid/high end cards on one process with transistor counts of 1/2/4 billion, then you'd expect the next gen new process parts to be 2/4/8 billion transistors.
20nm high end will be between 10-14billion transistors. Usually you get a 50% power reduction, so twice the transistors = same power usage as previous gen. 20nm this time around is said to only be 20% power reduction which is the key problem with it. So we could see 14billion transistors, 80% faster high end, but then it's likely to use 400W. Instead we might see something closer to 30-40% faster, 40-50% higher transistor count and 250-280W.
I've seen a couple rumours that the power reduction was improved towards 30% which would make things a little easier, but 20nm looks a bit of a wash.
We could see what is effectively the current high end shrunk to a midrange part, using about 20% less power. 10% bump in transistor count and a little more performance 10-15% over a 780ti looks sensible for the next gen midrange part. The suggested TDP there pretty much points at the problem of terrible power reduction on 20nm compared to previous new processes. You need both 2x transistor density(or close) and very importantly 50% or more power reduction to achieve double transistor count. Take a current 780ti and make it twice the transistor count and it would be 500W, * 50% power reduction = 250W, 20% power reduction and it's 400W. 250W with a 20% power reduction is 200W, with 10% higher transistor count would be 220W, and the final 10W coming from very high clocked memory to give almost enough bandwidth on a narrow bus and we have an underwhelming card. With 256bit bus you'd expect something that had some serious horsepower for 1080p but be pretty woeful at higher resolutions. Certainly not looking at the first gen of really good at 4k cards there.
This is the exact part I've been saying we'll see for the past year.
AMD/Nvidia at best doing effectively a high end shrink to a midrange part on the next process. 7970/680gtx replacements on 20nm with 290/780 +10-15% performance. 16nm finfet designs a year later for 290/780 replacements with +80% performance.
Comparison of 16nm FinFET 7.5T to 20nm 9T: 37% less power at the same speed, or 26% better performance at the same leakage
Quote on TSMC's process, this is fundamentally why the 16nm process will deliver the 290/780ti replacements. So a 2x 780ti(transistor count and power) would be around 400W on 20nm, but 240W on 16nm. it's really going to make the 7970/680gtx replacements boring. Nvidia is likely to pretty much charge the same as a 780ti for the new card with it being maybe 20% faster at low res, maybe slower at high res) but the performance difference won't be mind blowing. AMD will be cheaper but £350-400 for a card barely faster than 290x just doesn't scream excitement.
Obviously if you have an older card the new cards will be better to upgrade to than a 290/780ti, but if you have a high end card, I think late 2015 for real high end replacements will be the first opportunity for a real game changing level of performance.
