@ballsleon said:
@korosuzo: hypothetically, or new similarly priced cars may outperform the 980 Ti, but... it's a crap shoot till we have more information.
Ye, although there have been scant few generational shift where the new GPUs have arrived on the market that give less performance for the same price. The real question is does the next generation big Pascal come out early (we know they're mass producing the huge chip they'd likely use for such a design, we don't know when they'll be releasing them to consumers) and, probably more importantly, does the smaller design that slots in below it actually end up faster than the old king-sized Maxwell chip. It's not completely unimaginable that the GTX980 (non-Ti, completely different chip) and derivative GTX970 get replaced by a new Pascal design that comes in around the same launch price but both healthily outperform the old 980 Ti. I'm not convinced by that (that even the GTX1070 will be clean sailing over the 980Ti for about $350-400 launch price), partially because I'm not sure what the memory controller is going to be at that $350-550 launch price point** cards. GDDR5 is possibly ok with a wide bus (something where nVidia have plenty of room to expand vs AMD's wide designs) but you'd really like to see them offering HBM2 (which is what the high end Pascal chip uses to get massive bandwidth). Mainly it's because we're pretty confident a GP104 could be made to blow the old GM200 out of the water (see the thread on Pascal and those GFLOPS numbers for GP100: that's a big jump even for FP32 [where most shader code currently runs]) but does nVidia feel they need to (to compete with AMD's Polaris cards - competition is good!) or are they looking to get away with a cheaper second tier (smaller chip) that's a lot further behind the big chip than the GM204 was behind the GM200.
Sometimes we get old GPUs being flushed from inventory with good deals but often buying into the new generation is a good plan (new features in the architecture sometimes matter*, less Watts and so less heat and less noise and less PSU load for the same perf from architecture and die shrinks). At this point we expect someone to announce something within months so if you can wait, now is an excellent time to wait. AMD are also making noises about Polaris (their new design) with windows of the Summer for a full release of the various models so that can't be that far away. As soon as we see the first consumer model announced (we know about the big Pascal chip because of the announcement of it for use in servers, there is no consumer model yet), we should have a much better idea of what sort of performance gain at that price point the companies are targeting (and so can more confidently extrapolate likely other releases).
* eg Pascal chips can often merge two (floating-point: the maths almost all GPU work is based on) FP16 [16-bit] operations into a FP32 [32-bit] unit so any shader maths using FP16 will be up to twice as fast for the same float unit [shader unit] count; Maxwell 2+ gives programmable sampling so the anti-aliasing can be slightly better due to temporal dithering and some games can exploit this to make nice AA but it won't work on earlier chips; GCN 1.1+ has a nice 3D audio hardware block; all the modern designs [Kepler+, GCN 1.0+] can do almost-free video recording as you game.
** Reminder: the GTX980 (non-Ti) launched for $550, even if today you can get the much larger 980 Ti for about $600 and the 980 has come down in price to adapt to that.
Log in to comment