- Intel benefits from mobile growth even if it's not selling the phone hardware. Each of those phones needs to connect to back-end servers for almost everything, and those servers run X86. Every time they connect to a server, it generates data, and that data is analyzed and monetized by servers that, again, run X86.
- Fabs, and by proxy fab time, is super expensive (and growing superlinearly as process size shrinks). They may have made a strategic decision that they'd rather be pumping out Xeons for backend servers with that capital than Atoms for phone clients, except to the extent that they design and produce enough to "stay in the game" should they wish to change direction.
- Judging from the rate of improvement of the Intel low-power CPUs and GPUs, they are on a relatively short path to being extremely competitive with ARM and the associated graphics chipsets.
- Intel has a lot of experience writing "sufficiently smart compilers", and doing CPU / GPU integration. Both of those are extremely handy for low-power graphics-heavy environments.
I think this is just one of those things where it doesn't look like a trend until it's inevitable.
Interestingly, Intel spends a lot of die budget on on-chip instruction level parallelism, and their major instruction set-level enhancements have been increasing vector operation lengths. They've also invested somewhat heavily in GPGPU-style vector coprocessors in their Xeon Phi's.
Right. But the whole point of EPIC was to do away with that part of the die budget entirely, and a lot more besides.
GPGPU programming has worked because nobody proposed to run general code on it from the beginning. It just so happens that there are some workloads that fit the vector processing model.
- Intel benefits from mobile growth even if it's not selling the phone hardware. Each of those phones needs to connect to back-end servers for almost everything, and those servers run X86. Every time they connect to a server, it generates data, and that data is analyzed and monetized by servers that, again, run X86.
- Fabs, and by proxy fab time, is super expensive (and growing superlinearly as process size shrinks). They may have made a strategic decision that they'd rather be pumping out Xeons for backend servers with that capital than Atoms for phone clients, except to the extent that they design and produce enough to "stay in the game" should they wish to change direction.
- Judging from the rate of improvement of the Intel low-power CPUs and GPUs, they are on a relatively short path to being extremely competitive with ARM and the associated graphics chipsets.
- Intel has a lot of experience writing "sufficiently smart compilers", and doing CPU / GPU integration. Both of those are extremely handy for low-power graphics-heavy environments.
I think this is just one of those things where it doesn't look like a trend until it's inevitable.