was it everything you had hoped for? i think three good things about it are the low price, the standard of 4GB of GDDR5, and the fact that dp was not as crippled as nvidia's best offerings are. and also the performance, even though the 7970 GHz edition would be more than fast enough for me.
id say the things i like least about it are the lack of having good texture units (still only half speed 64 bit textures and the trilinear and lod calcs are still probably not as good as those of nvidia plus the fact that nvidia's af has just the right amount of angle variance), the high temps (at least on reference models, although nvidia isnt all that great at load temps either), the fact that it isnt really a brand new architecture, and the fact that they reduced dp performance from 1/4 to 1/8.
i dont know much about mantle, so im not commenting on that now.
i hope maxwell will run at lower temps than kepler or at least not do away with temp target. i hope that maxwell's drivers will not be limit it in dp performance, i hope that it will be backwards compatible with existing cuda apps, i hope they dont decrease texture filtering quality in any way... although maybe they could add a new trilinear filtering algorithm that looks even better. and finally, i hope they carry every single one of kepler's display logic features over to maxwell's display logic.
none of the follow will going to happen, but i also wish they would:
get rid of render target and color buffer modes less than 64 bit, get rid of 24 and 16 bit z-buffer formats (probably have only D32FX and D32FPS8X24 formats if not D64FX and D64FPS16X48, although 32 bit log z buffers look pretty damn sweet so 64 bits may not be necessary for the depth buffer), make the ROPs do 64 bit render targets at full speed (and 128 bit ones at half speed), doing away with dx compliance (or at least make mostly beta drivers so they dont have to comply with microsoft's demands) with dx apps while just using a wrapper to opengl or simply their own api.
if they ditched dx-compliance, they could make it so programming would be more like programming to the metal. dx is for amature programmers while open gl brings out the best results from the best 1% of programmers.
maybe i lack imagination, but GK110 is actually perfect except for the drivers which are still decent and they can always enable full DP for the 780GTX and 780GTX Ti for specific games down the road because double precision will eventually be standard.
id say the things i like least about it are the lack of having good texture units (still only half speed 64 bit textures and the trilinear and lod calcs are still probably not as good as those of nvidia plus the fact that nvidia's af has just the right amount of angle variance), the high temps (at least on reference models, although nvidia isnt all that great at load temps either), the fact that it isnt really a brand new architecture, and the fact that they reduced dp performance from 1/4 to 1/8.
i dont know much about mantle, so im not commenting on that now.
i hope maxwell will run at lower temps than kepler or at least not do away with temp target. i hope that maxwell's drivers will not be limit it in dp performance, i hope that it will be backwards compatible with existing cuda apps, i hope they dont decrease texture filtering quality in any way... although maybe they could add a new trilinear filtering algorithm that looks even better. and finally, i hope they carry every single one of kepler's display logic features over to maxwell's display logic.
none of the follow will going to happen, but i also wish they would:
get rid of render target and color buffer modes less than 64 bit, get rid of 24 and 16 bit z-buffer formats (probably have only D32FX and D32FPS8X24 formats if not D64FX and D64FPS16X48, although 32 bit log z buffers look pretty damn sweet so 64 bits may not be necessary for the depth buffer), make the ROPs do 64 bit render targets at full speed (and 128 bit ones at half speed), doing away with dx compliance (or at least make mostly beta drivers so they dont have to comply with microsoft's demands) with dx apps while just using a wrapper to opengl or simply their own api.
if they ditched dx-compliance, they could make it so programming would be more like programming to the metal. dx is for amature programmers while open gl brings out the best results from the best 1% of programmers.
maybe i lack imagination, but GK110 is actually perfect except for the drivers which are still decent and they can always enable full DP for the 780GTX and 780GTX Ti for specific games down the road because double precision will eventually be standard.
Facebook
Twitter