All important specs are on the slide. According to the power connectors, the maximal power draw of the boards is 300W. I hope that Mantle will not drain all the energy of AMD driver team because it would be nice to have an OpenGL 4.4 support in Catalyst…
By me, AMD Radeon R9 290X has a faster memory bus than NVIDIA GTX Titan. So, in theory, R9 290X should be faster. However if specification of NVIDA Tesla K40 (Atlas) will be true, it may be very difficult war, because K40 will have a slower memory bus ( 384-bit ), but it will have a bigger memory size (12 GB). In my point of view, NVIDIA K40 won’t use all memory, because memory bus is too slow for this memory size.
For crunching, size of shared memory and private register space is more important and number of ALUs(assuming opencl/cuda code is properly optimized). For gaming, number of rops, number of tmus, global memory bandwidth and drivers.