Kernel Optimization

[Guest Winston Ma]

Hi all,

I read the benchmark score (Quadrant and Linpack) of Nexus One and Dell Streak, which both use the same CPU. I wonder why they perform 20-30% better on the CPU score, under the same clock. It seems to me that the CPU is running more efficient on their devices. I wonder if we could "merge" the code from their kernel tree to ours to improve the CPU efficiency.

Please share your thoughts everyone.

Edited April 7, 2011 by Winston Ma

[Guest zorginho]

Use a custom like miui 1.3.5 or cm7 or leoginger 3.20, the scores and the general speed is really better than stock/stock based roms.

Edited April 7, 2011 by zorginho

[Guest Auxx]

Use T&L with 1600 patch - it will give you insane test results.

[Guest zorginho]

Use T&L with 1600 patch - it will give you insane test results.

but that patch is only a cheat... use a rom with a "real optimised kernel" then we can speak again!!

[Guest Winston Ma]

but that patch is only a cheat... use a rom with a "real optimised kernel" then we can speak again!!

Yes agree. That's why I point out both Quadrant and Linpack score. In both cases Nexus One got 20-30% higher CPU score.

The only difference is N1 use 2.6.32 kernel but I don't think 20-30% could come from newer kernel. As they use the same CPU we could copy their CPU code (assembly?) and tried that on our device. Honestly I never compile the kernel for Android. I would love to try.

[Guest Dr. Robert Chase]

benchmarks = bullshit.

[Guest Dario93]

benchmarks = bullshit.

This sentence reminds me Varyag again :D

[Guest Delnar_Ersike]

This sentence reminds me Varyag again :D

quadrant = bullshit

dont trust it

Pretty much down to the letter. B)

Benchmarks are pretty complicated as far as scores go, and they rarely mirror actual performance. For example, the Quadrant score I got when using T&L .14 with the quadrant patch is far greater than the one I'm currently getting with LeoGingerbread 3.20, even though I've experienced much smoother performance and faster app load times in LeoGingerbread. LinPack is pretty much the same story: I remember when I tested a version of LCR (I think it was 1.8) against a version of Xian's Eclair ROM, and although Xian's ROM had a slightly lower LinPack score than LCR (I think there was a 1 or 2 point difference, this was when LinPack scores weren't in the 20's), high-end games ran much better under Xian's ROM than under LCR.

So just because N1 and Dell Streak scores are slightly higher than A1 scores, doesn't mean they actually run faster or use kernel optimizations that speed up performance.

So far, the most accurate benchmarks I found (the ones whose scores portray actual ROM performance the most accurately) are the SmartBench benchmarks (2010 and 2011) and NenaMark1, and in all three benchmarks, the scores I've received with my device overclocked to 998Mhz are either roughly equal to or slightly greater than the average scores of N1 users.

Edited April 7, 2011 by Delnar_Ersike

[Guest Winston Ma]

This sentence reminds me Varyag again :D

Pretty much down to the letter. B)

Benchmarks are pretty complicated as far as scores go, and they rarely mirror actual performance. For example, the Quadrant score I got when using T&L .14 with the quadrant patch is far greater than the one I'm currently getting with LeoGingerbread 3.20, even though I've experienced much smoother performance and faster app load times in LeoGingerbread. LinPack is pretty much the same story: I remember when I tested a version of LCR (I think it was 1.8) against a version of Xian's Eclair ROM, and although Xian's ROM had a slightly lower LinPack score than LCR (I think there was a 1 or 2 point difference, this was when LinPack scores weren't in the 20's), high-end games ran much better under Xian's ROM than under LCR.

So just because N1 and Dell Streak scores are slightly higher than A1 scores, doesn't mean they actually run faster or use kernel optimizations that speed up performance.

So far, the most accurate benchmarks I found (the ones whose scores portray actual ROM performance the most accurately) are the SmartBench benchmarks (2010 and 2011) and NenaMark1, and in all three benchmarks, the scores I've received with my device overclocked to 998Mhz are either roughly equal to or slightly greater than the average scores of N1 users.

Thanks everyone for the discussion. In the following discussion I would like to focus on CPU score only.

I agree that Quadrant CPU score is a combination of several portion like A/V decoding which would make the score different. However I believe linpack is a relatively CPU-only testing platform as they do nothing but just pure calculation. Therefore no doubt I have bias on linpack score.

I ran the smartbench score and I found the follow results:

Liquid E 2.2.2 P1158 G757

N1 2.2.2 P1146 G601

Dell Streak 2.2.2 P1170 G860

However Smartbench score is based on the overall experience on either Productivity or Gaming. It's a overall score between CPU and other parts. Yes I agree Liquid E is running great, the overall experience is good (Thanks the community). The point I want to make here is... is there room for us to improve the CPU efficiency.

What I believe is

1. There is room for Liquid E to optimize the CPU portion.

2. If the CPU code can be optimized, then the Smartbench score would be higher than the current score.

Edited April 7, 2011 by Winston Ma

[Guest xdxdxdxdxdx]

quadrant score boost up is the result off turn on the stagefright(acer turn it off on stock rom)

Edited April 8, 2011 by xdxdxdxdxdx