ARM floating point performance

[codedivine]

Hi folks. I wrote an article for Anandtech about floating point performance on some ARM cores. http://www.anandtech.com/show/6971/exploring-the-floating-point-performance-of-modern-arm-processors

Thought you would be interested

 

[willardjuice]

I read it last night and can vouch it's well worth checking out!

Get Anand to give you an ipad 4 and test swift!

 

[Nebuchadnezzar]

Would you publish the benchmark for more testing?

Still interested in Krait 300 vs A15 comparisons. Your own RgbenchMM shows quite a difference on the two S4 variants: 3800MFlops max on the 1.9GHz S600 vs 4700MFlops on the 1.6GHz Octa.

 

[Laurent06]

Hi folks. I wrote an article for Anandtech about floating point performance on some ARM cores. http://www.anandtech.com/show/6971/exploring-the-floating-point-performance-of-modern-arm-processors

Thought you would be interested

Already read it and liked it, thanks!

I will second Nebuchadnezzar question and would like to ask if it'd be possible to also get latency information on top of throughput.

 

[codedivine]

Would you publish the benchmark for more testing?

Still interested in Krait 300 vs A15 comparisons. Your own RgbenchMM shows quite a difference on the two S4 variants: 3800MFlops max on the 1.9GHz S600 vs 4700MFlops on the 1.6GHz Octa.

Yes I will put it up somewhere soon.

As for RgBenchMM, I have been doing some experiments on Krait at least and believe the issue is memory access and not FP performance. Inserting a few prefetch instructions explicitly improves performance substantially.

 

[Helmore]

As for RgBenchMM, I have been doing some experiments on Krait at least and believe the issue is memory access and not FP performance. Inserting a few prefetch instructions explicitly improves performance substantially.

This is precisely why you started that article, right? To figure these kinds of things out so that coders can better optimize their code, if they deem it necessary.

Would you say that this is a disadvantage for Krait compared to Eagle (Cortex-A15)? I'm a bit of a noob when it comes to the finer details of chip architecture, but would Qualcomm be able to improve this by improving their prefetcher? Another question is of course whether that's worth it, but that's a question that Qualcomm has to answer for themselves.

 

[codedivine]

This is precisely why you started that article, right? To figure these kinds of things out so that coders can better optimize their code, if they deem it necessary.

Yes :smile:
Trying to write RgBenchMM was not a great experience. I was essentially shooting in the dark while trying to improve the code, as far as official documentation went.

Would you say that this is a disadvantage for Krait compared to Eagle (Cortex-A15)? I'm a bit of a noob when it comes to the finer details of chip architecture, but would Qualcomm be able to improve this by improving their prefetcher? Another question is of course whether that's worth it, but that's a question that Qualcomm has to answer for themselves.

Well not very impressed with the prefetcher in Krait so far, but my experience is also fairly limited till now. I have not yet experimented with Cortex A15 in the memory department. You can expect to see an article about that sometime perhaps, though it will take me a while (1-2 months) depending on how much work I need to do at univ (I am a grad student).