Hacker News Reader

Comment by berti | Hacker News Reader

Specint benchmarks conducted by Anandtech show the A13 going toe-to-toe with the 9900K [0], although x86 still has a decent lead for floating point.

[0] https://www.anandtech.com/show/14892/the-apple-iphone-11-pro... (scroll down to the 2nd last plot, look at the right half)

View on HN

Do mobile chips focus less on floating point for some reason? Could that explain the seemingly massive difference in perf/watt?

Floating point tends to matter for classic 'number crunching' workloads- crypto, media encoding, anything to do with modeling the physical world, image processing, AI.

Phones tend towards 'control flow', e.g. display an interface, accept a user input, communicate with a server, display information... and media decoding & gps both have dedicated hardware.

FP workloads can be incredibly power hungry, so it's fortunate mobile devices mostly manage to avoid them. But any chip with an FP unit is going to shut it off when running integer code.

Crypto doesn't do floating point. Media transcoding does SIMD, but not usually not floating point.

Fair enough wrt media. However SIMD e.g. MMX is done in the floating point unit (https://softpixel.com/~cwright/programming/simd/mmx.php)

I thought bitcoin mining was floating point, and that was why GPUs were so adept.

MMX is obsolete.

SSE/AVX has its own set of registers. It can do floating point and integer operations. It would not be correct to refer to it as a floating point unit.

GPUs are good at mining because they can do so many calculations in parallel. They're just as capable of integer operations as the SIMD units on a CPU.

MMX was integer SIMD that re-used the FPU register file, but was not doing floating-point operations.

Crypto stuff never uses floating point. It might have built-in instructions like AES-NI, though. In the test geekbench consider that - integers ones[0] and definitely not floating point. Media encoding tend not to use floating point, either. (decoders tend to be built-in in the GPUs with designated hardware, so testing on CPU is quite pointless)

[0]: https://www.geekbench.com/doc/geekbench5-cpu-workloads.pdf

Reply on news.ycombinator.com