Anandtech has a great piece on the Imagination Technologies "Rogue" graphics IP (used in iPhone 5S, iPad retina Mini and iPad Air), including the next generation announced 6XT cores. They're only able to cover the core shader units (Unified Shading Clusters as ImgTec calls them), but it's a great look at their arithmetic capabilities and why you can't compare them directly to Nvidia's Tegra K1 "cores".
In short, Rogue pipelines have wide FP16 (floating point 16 bit) ALUs (Arithmetic Logic Units), which allows them to do many FP16 operations in one cycle. Desktop GPUs have done away with FP16 ALUs all together and usually have 2 FP32 ALUs per pipeline (As does Rogue). This obviously saves space since they don't waste space with FP16 units, but ImgTec has determined that mobile workloads are better suited to using FP16 ALUs, which right away would make them more efficient since they're using less power to do a FP16 operation (which would waste power on a FP32 unit). In fact, ImgTec made it so that 6XT can execute even more FP16 operations per cycle with respect to 6, so they must really believe in this solution. 6XT can do four FP16 operations per cycle, in comparison to just the two FP32 operations it can do, along with its competitors.
So, if you have an even number of "cores" with a Rogue design vs. a K1 design, you're probably talking about more actual throughput for a Rogue design in a mobile environment due to its ability to focus on those FP16 operations.
This ignores other parts of the GPU, including ImgTec's Tile-Based Deferred Rending (TBDR) approach to GPU workloads, which is in itself more efficient than competing strategies or non-deferred renderers.
Basically, ImgTec doesn't want a core-war with Nvidia since they think their cores are better in a mobile environment. Nvidia has unified their mobile and desktop IP with K1 (Kepler, Maxwell and beyond) to simplify their design process and provide a uniform feature set across all their products. It makes sense for them but makes it difficult to be equal with a company that focuses only on mobile-optimized designs such as ImgTec does.
ImgTec has published two whitepapers as well. One is on their take on the core wars. The second details their brand new flagship part, the GX6650, which equals the K1 in core count but will far outstrip it in real world mobile workload scenarios because of its wider FP16 design and its other architecture efficiencies like the use of TBDR.
Look for the 6XT series to definitely be included in A-series Apple chip by A9, with some possibility of inclusion in A8. This is, of course, assuming Apple doesn't go full custom GPU (probably with ImgTec IP) with their Orlando GPU team.
In short, Rogue pipelines have wide FP16 (floating point 16 bit) ALUs (Arithmetic Logic Units), which allows them to do many FP16 operations in one cycle. Desktop GPUs have done away with FP16 ALUs all together and usually have 2 FP32 ALUs per pipeline (As does Rogue). This obviously saves space since they don't waste space with FP16 units, but ImgTec has determined that mobile workloads are better suited to using FP16 ALUs, which right away would make them more efficient since they're using less power to do a FP16 operation (which would waste power on a FP32 unit). In fact, ImgTec made it so that 6XT can execute even more FP16 operations per cycle with respect to 6, so they must really believe in this solution. 6XT can do four FP16 operations per cycle, in comparison to just the two FP32 operations it can do, along with its competitors.
So, if you have an even number of "cores" with a Rogue design vs. a K1 design, you're probably talking about more actual throughput for a Rogue design in a mobile environment due to its ability to focus on those FP16 operations.
This ignores other parts of the GPU, including ImgTec's Tile-Based Deferred Rending (TBDR) approach to GPU workloads, which is in itself more efficient than competing strategies or non-deferred renderers.
Basically, ImgTec doesn't want a core-war with Nvidia since they think their cores are better in a mobile environment. Nvidia has unified their mobile and desktop IP with K1 (Kepler, Maxwell and beyond) to simplify their design process and provide a uniform feature set across all their products. It makes sense for them but makes it difficult to be equal with a company that focuses only on mobile-optimized designs such as ImgTec does.
ImgTec has published two whitepapers as well. One is on their take on the core wars. The second details their brand new flagship part, the GX6650, which equals the K1 in core count but will far outstrip it in real world mobile workload scenarios because of its wider FP16 design and its other architecture efficiencies like the use of TBDR.
Look for the 6XT series to definitely be included in A-series Apple chip by A9, with some possibility of inclusion in A8. This is, of course, assuming Apple doesn't go full custom GPU (probably with ImgTec IP) with their Orlando GPU team.