TR Forums

Samsung Vibrant (Hummingbird Cortex A8) - 205
Nexus 7 -
1 thread - 197
2 thread - 526
4 thread - 920

derFunkenstein wrote:

Samsung Vibrant (Hummingbird Cortex A8) - 205
Nexus 7 -
1 thread - 197
2 thread - 526
4 thread - 920

Thanks

Dear Gerbils. I am pretty sure after this post, many of you are going to get ready to kill me.

I have reworked the code slightly to give MUCH higher performance. Pushed v1.2 on the site. http://rgbench.com as always.
The previous version (1.1) is still a valid and correct benchmark, but this one now gives much higher performance on the same task.

Thanks for the update! It now runs on my CM7-hacked Motorola Triumph, which is very low-end compared to most phones on this thread:

Using version 1.2:

1. 1 Thread: 960 MFlops (this is a single-core CPU)
2. 2 Threads: 400 MFlops
3. 4 Threads: 381 MFlops

chuckula wrote:

Thanks for the update! It now runs on my CM7-hacked Motorola Triumph, which is very low-end compared to most phones on this thread:

Using version 1.2:

1. 1 Thread: 960 MFlops (this is a single-core CPU)
2. 2 Threads: 400 MFlops
3. 4 Threads: 381 MFlops

960 MFlops on a single-core Snapdragon S2
Not what I was expecting. I would have expected more like 400 MFlops or so. Are you sure that is not a typo?

For everyone's benefit, scores collected from other users :

1. Nexus 7 : 1470 MFlops
2. Galaxy S2 (international, Exynos 4410): 890 MFlops
3. Galaxy S2X (T-mobile, Snapdragon S3): 1040 MFlops
4. Transformer TF201 (Tegra 3, 1.3GHz): 1881 MFlops
5. Galaxy S3 (international Exynos 4 Quad): 2189 Mflops

codedivine wrote:

chuckula wrote:

Thanks for the update! It now runs on my CM7-hacked Motorola Triumph, which is very low-end compared to most phones on this thread:

Using version 1.2:

1. 1 Thread: 960 MFlops (this is a single-core CPU)
2. 2 Threads: 400 MFlops
3. 4 Threads: 381 MFlops

960 MFlops on a single-core Snapdragon S2
Not what I was expecting. I would have expected more like 400 MFlops or so. Are you sure that is not a typo?

The first results I submitted were what the phone displayed, but I just ran the benchmark again and I got different results. I changed my settings so that the phone does not turn off the screen or go to sleep during the benchmark run. This is a CM-7 phone with Android 2.3.7, but the CPU is not overclocked (1 GHz base clock).

Here's another run using 1 thread: 335 MFlops.

So there ya go.. sometimes there might be an outlier sneaking into the mix

chuckula wrote:

The first results I submitted were what the phone displayed, but I just ran the benchmark again and I got different results. I changed my settings so that the phone does not turn off the screen or go to sleep during the benchmark run. This is a CM-7 phone with Android 2.3.7, but the CPU is not overclocked (1 GHz base clock).
Here's another run using 1 thread: 335 MFlops.
So there ya go.. sometimes there might be an outlier sneaking into the mix

Thanks for the update. Sounds more reasonable now.
Hoping some TR and other folks will pick the benchmark up. I think it is now ready for prime-time

From a user with a Snapdragon S4 based HTC One X:

Single thread: 730 MFlops
Two threads: 1460 MFlops

Clearly, Krait is showing a much higher performance than either Qualcomm S3 or the ARM Cortex A9
The quad-core S4 is going to be a beast

Now live in Google Play: https://play.google.com/store/apps/deta ... ne.rgbench

codedivine wrote:

Clearly, Krait is showing a much higher performance than either Qualcomm S3 or the ARM Cortex A9

As expected. Have you any plans to make an iOS compatible version to test the A6?

codedivine wrote:

Now live in Google Play: https://play.google.com/store/apps/deta ... ne.rgbench

I pushed it to my Nexus 7 and my wife's Acer Tegra 2 tablet and I'll just verify that I see similar performance when I get home.

codedivine wrote:

For everyone's benefit, scores collected from other users :

1. Nexus 7 : 1470 MFlops
2. Galaxy S2 (international, Exynos 4410): 890 MFlops
3. Galaxy S2X (T-mobile, Snapdragon S3): 1040 MFlops
4. Transformer TF201 (Tegra 3, 1.3GHz): 1881 MFlops
5. Galaxy S3 (international Exynos 4 Quad): 2189 Mflops

codedivine wrote:

From a user with a Snapdragon S4 based HTC One X:

Single thread: 730 MFlops
Two threads: 1460 MFlops

TI OMAP 4460 Scores: Underclocked to 1GHz from 1.2GHz
Single thread: 295.0 MFlops
Two threads: 778.0 MFlops

derFunkenstein wrote:

codedivine wrote:

Now live in Google Play: https://play.google.com/store/apps/deta ... ne.rgbench

I pushed it to my Nexus 7 and my wife's Acer Tegra 2 tablet and I'll just verify that I see similar performance when I get home.

You will see *slightly* different performance. Unfortunately, even in v1.2 there was a bug. (I wanted to die of shame, again, but here I am still alive )
However, everything looks good now.

DancinJack wrote:

TI OMAP 4460 Scores: Underclocked to 1GHz from 1.2GHz
Single thread: 295.0 MFlops
Two threads: 778.0 MFlops

This is from the Play version ?

codedivine wrote:

You will see *slightly* different performance. Unfortunately, even in v1.2 there was a bug. (I wanted to die of shame, again, but here I am still alive )
However, everything looks good now.

Please - keep on living. You're way up on me, I wouldn't even know where to start with such an app. Or any app for that matter.

codedivine wrote:

This is from the Play version ?

Indeed. I'll bump it back up to 1.2 and see how it goes.

I should include that the kernel I'm using has a tweaked governor. Throttling works a little differently. The one I use makes you utilize a higher percentage of CPU before it jumps to the next speed up. This is all in an attempt to conserve battery. I'll test with the one I use and stock OMAP kernel governor.

My regular settings: 230MHz min CPU - 1036MHz max CPU - dynamic interactive gov.
Single thread: 295.0 MFlops
Two threads: 778.0 MFlops

Stock OMAP 4460 speeds: 350MHz min CPU - 1200MHz max CPU - dynamic interactive gov.
Single thread: 340.0 MFlops
Two threads: 900.0 MFlops

Stock OMAP 4460 speeds and stock OMAP gov. (350-1200-ondemand)
Single thread: 340.0 MFlops
Two threads: 910.0 MFlops

Thanks. Here are my results from the Play version (Galaxy S2X using Snapdragon S3 dual-core 1.5GHz):

1 thread: 495 MFlops
2 threads: 1175 MFlops

The lowly OMAP in my Nexus weeps. At least I get updates and stock Android?

I like that it has a progress indicator now.

It would be good to disable the onscreen controls (the dropdown list and the Run button) until the test is completed. Either that or use wakelock to keep the CPU from entering sleep state.

Also, it should be possible to keep the screen from sleeping while the benchmark is actively running.


Galaxy Note, Snapdragon S3 @ 1.5GHz, running Android 4.1 (JB, CM10), RgbenchMM 1.3
2 threads: 1153 MFlops
1 thread: 572 MFlops

ChronoReverse wrote:

I like that it has a progress indicator now.
It would be good to disable the onscreen controls (the dropdown list and the Run button) until a test it complete though.
Also, it should be possible to keep the screen from sleeping while the benchmark is actively running.

Thanks for the suggestions. Now that the benchmark actually works, I can look into doing this.

Galaxy Note, Snapdragon S3 @ 1.5GHz, running Android 4.1 (JB, CM10), RgbenchMM 1.3
2 threads: 1153 MFlops
1 thread: 572 MFlops

Thanks.

Don't forget to leave a rating everyone

DancinJack wrote:

The lowly OMAP in my Nexus weeps. At least I get updates and stock Android?

Hmm, considering it's an A9 chip that should have higher per clock performance than the Snapdragon S3 (I'd say at least 10%), the settings for your CPU are hindering its performance significantly. I suspect the undervolting might be too aggressive for instance. The low single thread scores also indicate an overly aggressive governor.

At 1.2GHz (1188MHz actually), my Snapdragon S3 gets 959 MFlops for 2 threads and 476MFlops for a single thread.

codedivine wrote:

Thanks for the suggestions. Now that the benchmark actually works, I can look into doing this.

Heh, I'll just keep "demanding" new features then

A history of past results (and perhaps a running average) would be nice (especially if the different number of core results are kept separate).

Droid 3, rooted with AOKP 4.0.4, with an OMAP 4430 @ 1GHz, running the version from Google Play.

1 Thread - 302.0 MFlops.

2 Threads - 808.0 Mflops.

I'm using the mot_hotplug CPU scaling governor, but I think that was the default choice. I may post back later with results from the performance scaling governor and some of the others.

Ondemand scaling governor with 1 thread was 298.0 MFlops.
Performance with 1 thread was 302.0 MFlops.

ChronoReverse wrote:

Hmm, considering it's an A9 chip that should have higher per clock performance than the Snapdragon S3 (I'd say at least 10%), the settings for your CPU are hindering its performance significantly. I suspect the undervolting might be too aggressive for instance. The low single thread scores also indicate an overly aggressive governor.

It's actually not undervolted at all. The governor is the primary cause for the lowered performance I suspect. Honestly it's perfectly fine with me. My phone never feels bogged down or sluggish. I can now get over 24hrs on a single charge with LTE on too because of the tweaks I've made. I'm happy with performance and battery life at this point.

ChronoReverse wrote:

I like that it has a progress indicator now.

It would be good to disable the onscreen controls (the dropdown list and the Run button) until the test is completed. Either that or use wakelock to keep the CPU from entering sleep state.

Also, it should be possible to keep the screen from sleeping while the benchmark is actively running.

Done both. Pushed v1.3.1 on Play (otherwise app is unchanged).

edit: Looks like Play store takes a little bit to actually publish update. Anyway, should be live soon.

Work has begun on the next app: An app to test ARM NEON performance

I got 1.3.1 on my Nexus 7 and my wife's A100.

The Nexus 7 results aren't really any different. I got 292 1 thread and 1488 4 thread.

The Acer A100 (Tegra 2) got 296 1 thread and 714 for 2

Still weird to see the huge 1-thread drop-off, but the results are consistent.

derFunkenstein wrote:

I got 1.3.1 on my Nexus 7 and my wife's A100.

The Nexus 7 results aren't really any different. I got 292 1 thread and 1488 4 thread.

The Acer A100 (Tegra 2) got 296 1 thread and 714 for 2

Still weird to see the huge 1-thread drop-off, but the results are consistent.

Thanks. That huge dropoff in single thread performance on the Cortex A9 is certainly puzzling. I have a Cortex A9 linux board (Pandaboard) and I will look into collecting more indepth data about what the hardware is doing.

EDIT: Actually, not so sure now. Looking.

**************
I think there is some explaination for the fact that the Snapdragon S3 have better single threaded performance than Cortex A9 in this case.

I looked at generate assembly. In my innermost loop, there are 6 loads and 8 MAC VFP instructions. Turns out the load instructions are actually VFP instructions as well. From documentation gleaned from third parties, it looks like Snapdragon S3 VFP can actually issue one VFP load and VFP ALU operation in parallel, while Cortex A9 VFP cannot.

So weep Cortex A9 owners, weep