Ubuntu 12.04 ARM Performance Becomes Very Compelling | Hacker News Reader

Ubuntu 12.04 ARM Performance Becomes Very Compelling

phoronix.com

65 points6ren13 comments

A lot of applications used on ARM may still be crippled due to lack of NEON assembly code. Many number-crunching applications, like compression, have lots of MMX/SSE SIMD on x86, but no NEON, and will likely be slow on ARM until they do.

What do you mean? All dual core ARM chips and above (except Tegra 2) have NEON. Also all 1 Ghz and above Snapdragon chips.

Most smartphones now are armv7 and have been for the last 2 years or so.

I don't see an issue with a baseline being set for armv7+vfp+neon being a minimum processor feature set going forward for mobile devices (phones & tablets)

Well the hardfloat stuff was clearly worthwhile, but some meaningful benchmarks against x86 would be more useful. My ARM machines are all faster than my terrible Atom, but a lot slower than anything modern for cpu intensive benchmarks.

how long does it take to compile the linux kernel?

It takes "ages", but that is mainly due to the very slow mass storage speed that current ARM SoCs feature. Mobile phones and development boards like the Pandaboard use SD cards as mass storage and SD cards can be very slow when not used for typically linear read/write operations in your digital camera

I have an ssd on one of my machines, which helps but that is a slower softfloat single core box, so it still takes ages.

Do you have any data to back this up? I have compiled kernels on an ARMv5 device off a (I believe PCI-e-hosted) 7200rpm SATA hard disk, and it _is_ slow.

I have only anecdotal evidence from people that are masochistic enough to compile their stuff natively, I prefer to cross-build.

This particular set of benchmarks are misleading as more of the performance comes from a faster CPU clock speed. For a more accurate feel, look at the middle set of numbers.

That said, hardfp looks to be a definite gain over softfp and really should have been deployed by distributions when the Cortex-A8 generation of processors became available.

There are very few hardfp benchmarks that really compare just this feature and leave the rest of the system identical. E.g. comparing Debian armv5 with floating point emulation vs Ubuntu armv7hf of course yields a huge gain, but that mainly is not due to the fact that floating point values are now passed in fp registers.

If SW relies heavily on passing a handful of floating point values in function calls, it should be rewritten to use NEON/SIMD on ARM - and then all you pass is a pointer to the data, regardless of the calling convention. And once you move to writing the mission critical stuff in assembly, you can chose your own calling convention for subroutines as you see fit anyway.

For reference, this is the same CPU as the Galaxy Nexus, run at the same clockspeeds.

Cortex A15 and DDR3 for ARM should really start to make the ARM architecture very compelling for (low-end) laptops with very high battery life. And they're both arriving this year.

Reply on news.ycombinator.com