3 ms·
> The overall point is that work is being increasingly done _not_ on the CPU. Eh? GPGPU has been a thing for decades and yet barely made a dent in the demand f
by kllrnohj 2y ago
> The overall point is that work is being increasingly done _not_ on the CPU.
Eh? GPGPU has been a thing for decades and yet barely made a dent in the demand for CPUs. Heck, CUDA is 17 years old!
The world has not budged from being CPU-centric and it isn't showing any signs of doing so. GPUs remain an accelerator for specialized workloads and are going to continue to be just that.
- astrange 2y agoHmm, this isn't power-efficient thinking. iPhones have lots of accelerators beyond GPUs - not because they're more performant but because they're better at performance/power. Of course they also add GPU-like things to the CPU too for the same reason: https://developer.arm.com/documentation/109246/0100/SME-Overview/SME-and-SME2 https://developer.arm.com/documentation/109246/0100/SME-Over...
- kllrnohj 2y agoThat's orthogonal. Fixed-function hardware has a power efficiency advantage over programmable hardware. This isn't new or surprising, but it's also unrelated to CPU vs. GPU. Both CPU & GPU are programmable, so that's not relevant.
- astrange 2y agoThose aren't useful distinctions. You might as well say the ALUs in a CPU aren't programmable because they aren't the instruction decoders.
- kllrnohj 2y agoHuh? Weren't you referring to things like the video decoding blocks? Those are genuinely fixed-function things, they have no aspects of programmability to them. It's why their capabilities are fixed for all time (eg, AV1 support requires a hardware change). It's like a pre-shader GPU. It's a very useful distinction to make.