12GB is not overkill, only a year ago people were saying how a 3GB NVIDIA card was OK and AMD was overkill with 4GB.
TitanX is ultimate, you want to play Project cars at 12k you need 4x TitanX and 12GB VRAM.
TitanZ was overpriced, was it a failure? No, NVIDIA sold them all and have even had to manufacturer more for us, hence why Asus is back on pre-order as the world wants more stock, they are arriving soon.
What is overkill to one person is the perfect card for another.
12GB for a gaming card IS overkill, I don't think anyone said 3GB wasn't asking for trouble but ultimately when there were 3GB cards they were what 2-3 years from it really being a problem as it might have been in the past year.
Ultimately people simply buy whatever is deemed best in reviews with little knowledge. The same way people bought some low end whatever, 9300gt with 2GB memory with their Dell computer because the 4870 1GB looked slower.
Most people who wanted to spend £800 on a Titan would have done so if it had 12, 6 or 3gb of memory, that is just how people are. People buy what's available and often what is best.. the actual specs of what is deemed best isn't as important as the status of being what is perceived to be best for many people.
people bought Titan Z's because they read they were the best most expensive cards around, mostly people that had more money than sense, that is fine, but most of those would still have bought a Titan Z with less memory as well.
New technology is always great, don't want to be stuck in dark ages, am sure it will be great and future generations no doubt awesome, or it could be like that memory Intel did one time, was it Rambus, silly fast, but expensive, crazy hot and never took off, think it was on the Abit TH7-II if my memory serves me right.....
Memory consumes little power, so even though HBM consumes less, most of the power draw and heat comes from the GPU, this is where AMD are not improving upon at the same rate as NVIDIA, but I like a hot card, that means more fun overclocking it and cooling it, sets the men apart from the boys.
Yes and no, a rough comparison is the same amount of GDDR5 and HBM providing 512GB/s of bandwidth would run around 80W for GDDR5 and 30W for HBM, the majority of the difference isn't in the memory but the memory controller which is on die. Communication off package is very power intensive, it's why literally every chip made in mobile, by Intel, by AMD and by Nvidia has brought as much on package if not on die as possible. There is a LARGE power saving to be had by using HBM over GDDR5.
There is no bandwidth level GDDR5 can't achieve that HBM can, what can be realistically achieved and without a workable power level, GDDR5 and HBM are night and day.
Lets say 390X actually needs at least 512GB/s of memory. You'd be looking at a 512bit controller with 8Ghz clocks to achieve that, the die size for the memory controller and the power cost would leave significantly less power within the same power limit, to be used directly for the rest of the core. IE 300W power limit, 80W on GDDR5, 220W core, or 30W on HBM, 270W on the core. Smaller die, cheaper pcb, easier power delivery(look up the pictures of pascal, while a mock up that will be what can be done when effectively every power hungry chip is on package) are all bonuses on top of that but you're looking at HBM being able to directly provide an almost 25% power increase to the rest of the core. Think 15-20% higher clocks or better power efficiency is have 4000 shaders instead of say 3200, 25% more rops, tmu's, etc, etc.
HBM brings a large benefit and would be hugely responsible for how a 500-550mm^2 AMD chip could beat a 600mm^2 Nvidia chip without about the same power usage. More than that, we'll see, it wouldn't surprise me if AMD went without a lot of DP performance, but again that effective 25% increase in power budget could well mean they fit the compute performance in there. At that stage AMD would have a DP beast that in the professional market could wipe the floor with anything Nvidia can offer. Considering Hawaii did great with the likes of Apple and restablishing themselves in the professional market a killer even further ahead DP compute capable core with a level of bandwidth Nvidia can't come remotely close to offering(even more important in compute than high res gaming) AMD could make a killing.