Google put Moorefield in Nexus Play, Tegra K1 in Nexus 9

[Abwx]

I believe the results to be authentic, but what the benchmark is representative of is an entirely different matter.

To my understanding, Geekbench code is very simple and thus is more of a test of issue width and execution unit resources than anything else.

So it's no surprise that an in-order 7-wide machine does very well in this benchmark.

Actualy i m not questioning the score authenticity but the actual ST perf, this look like it s quite optimistic when looking at the numbers of other CPUs, i m quite skeptic that is has same IPC as a Core i3-3110M.

 

[Enigmoid]

I believe the results to be authentic, but what the benchmark is representative of is an entirely different matter.

To my understanding, Geekbench code is very simple and thus is more of a test of issue width and execution unit resources than anything else.

So it's no surprise that an in-order 7-wide machine does very well in this benchmark.

Geekbench consists of short tests making it ideal for tdp limited mobile chips. Just gave it a run on my laptop. Task manager reports that CPU usage is all over the place constantly spiking and dropping. Looking at HW monitor and CPU package power Geekbench 3 is generally between 12W and 39W on a 3630qm, mostly around 20-28W. A solid CB run for instance is around 37-38W. Its not very intensive at all.

3630qm @ 3.2 ghz

Actualy i m not questioning the score authenticity but the actual ST perf, this look like it s quite optimistic when looking at the numbers of other CPUs, i m quite skeptic that is has same IPC as a Core i3-3110M.

Geekbench 3 is weighted as a SB 2.5 ghz core = 2500 points.

http://support.primatelabs.com/kb/geekbench/interpreting-geekbench-3-scores

 

[III-V]

Geekbench 3 is weighted as a SB 2.5 ghz core = 2500 points.

http://support.primatelabs.com/kb/geekbench/interpreting-geekbench-3-scores

Very interesting. I did not know that.

 

[Khato]

I believe the results to be authentic, but what the benchmark is representative of is an entirely different matter.

To my understanding, Geekbench code is very simple and thus is more of a test of issue width and execution unit resources than anything else.

The reason why I don't pay any attention to ARM vs x86 using Geekbench as a metric is simple - DGEMM is a well-understood, relatively standard benchmark algorithm and yet Geekbench reports Haswell at under one tenth of what it's capable of. Take a look at any Haswell DGEMM result in Geekbench and you'll see around 1.3 GFLOPS/GHz... whereas OpenBLAS obtains around 13.6 GFLOPS/GHz.

Why trust any of the Geekbench results for x86 to be representative of actual performance capability when they can't get the basics right?

Note that I wouldn't be surprised if Geekbench's excuse for their poor x86 DGEMM showing is that they're still using x87 or something... but that's representative of the key issue with Geekbench - let x86 languish in legacy while providing full support for the latest and greatest with ARM.

 

[ams23]

Very impressive performance, esp. since it's in AArch32 mode rather than AArch64. Floating point should get a nice boost in AArch64 mode.

When Anandtech reviewed the iPhone 5s, moving from AArch32 test to AArch64 test in Geekbench 3, they measured an improvement of +25% in BlackScholes, +16% in Blur Filter, +195% in DGEMM, +119% in SFFT, +26% in N-Body, +51% in RayTrace ( http://www.anandtech.com/show/7335/the-iphone-5s-review/4 )

Here is a Geekbench 3 comparison between HTC Volantis Nexus 9 and iPhone 5s, both using AArch32: http://browser.primatelabs.com/geekbench3/compare/1014854?baseline=1002621

The single-core performance in this comparison is 80-90% higher than first gen Cyclone! Now, exactly how much improvement there will be from AArch32 to AArch64 with the Denver core is unknown at this time (the Cyclone core shows an improvement of ~ 35% overall). It is also unclear exactly what CPU clock operating frequency is being used for the Denver cores benchmarked here.

At Hot Chips, Denver was said to have a Geekbench 3 single-core score that was ~ 1.65x higher than R3 Cortex A15 in Tegra K1, and ~ 1.375x higher than Cyclone in A7 iPhone 5s. So I don't know exactly what to make of these two new AArch32 data points. The only way for these numbers to make sense is if NVIDIA showed AArch32 Geekbench 3 data for Denver at Hot Chips while comparing it to AArch64 Geekbench 3 data for Cyclone.

 

[Khato]

At Hot Chips, Denver was said to have a Geekbench 3 single-core score that was ~ 1.65x higher than R3 Cortex A15 in Tegra K1, and ~ 1.375x higher than Cyclone in A7 iPhone 5s. So I don't know exactly what to make of these two new AArch32 data points. The only way for these numbers to make sense is if NVIDIA showed AArch32 Geekbench 3 data for Denver at Hot Chips while comparing it to AArch64 Geekbench 3 data for Cyclone.

Not necessarily. Have to remember that the actual execution pipeline is running its own instruction set and hence could already be making use of such optimizations. For example, I highly doubt that Denver's encryption performance isn't making use of acceleration, especially when considered in relation to performance everywhere else. Whereas on the FP side it's 50/50 whether it's weaker by design or if it needs the FP vectorization of AArch64 to extract its performance capabilities.

 

[MisterLilBig]

Everyone ignoring that the Nexus 9 Denver based K1 SoC is running at 2.3GHz???
http://www.google.com/nexus/9/


Meaning that the score above is higher than on the Nexus 9.

 

[Nothingness]

Note that I wouldn't be surprised if Geekbench's excuse for their poor x86 DGEMM showing is that they're still using x87 or something...

That's easy to check. Just download and disassemble

 

[Nothingness]

When Anandtech reviewed the iPhone 5s, moving from AArch32 test to AArch64 test in Geekbench 3, they measured an improvement of +25% in BlackScholes, +16% in Blur Filter, +195% in DGEMM, +119% in SFFT, +26% in N-Body, +51% in RayTrace ( http://www.anandtech.com/show/7335/the-iphone-5s-review/4 )

Here is a Geekbench 3 comparison between HTC Volantis Nexus 9 and iPhone 5s, both using AArch32: http://browser.primatelabs.com/geekbench3/compare/1014854?baseline=1002621

The single-core performance in this comparison is 80-90% higher than first gen Cyclone! Now, exactly how much improvement there will be from AArch32 to AArch64 with the Denver core is unknown at this time (the Cyclone core shows an improvement of ~ 35% overall). It is also unclear exactly what CPU clock operating frequency is being used for the Denver cores benchmarked here.

At Hot Chips, Denver was said to have a Geekbench 3 single-core score that was ~ 1.65x higher than R3 Cortex A15 in Tegra K1, and ~ 1.375x higher than Cyclone in A7 iPhone 5s. So I don't know exactly what to make of these two new AArch32 data points. The only way for these numbers to make sense is if NVIDIA showed AArch32 Geekbench 3 data for Denver at Hot Chips while comparing it to AArch64 Geekbench 3 data for Cyclone.

The iOS score you show is not AArch32, it's ARMv7. There are big differences such as AES support.

 

[ams23]

Not necessarily. Have to remember that the actual execution pipeline is running its own instruction set and hence could already be making use of such optimizations. For example, I highly doubt that Denver's encryption performance isn't making use of acceleration, especially when considered in relation to performance everywhere else. Whereas on the FP side it's 50/50 whether it's weaker by design or if it needs the FP vectorization of AArch64 to extract its performance capabilities.

Nebu @ B3D confirmed that certain items such as integer crypto performance can be accelerated even on AArch32, and looking at the HTC Volantis Nexus 9 Geekbench integer data, that certainly looks to be the case.

That said, the floating point data should see a nice improvement when using AArch64, so there is still some upside in the Denver Geekbench 3 data beyond that of the AArch32 results.

 

[Nothingness]

Here is a real comparison against a real AArch32 build and CPU (on Cortex-A57): http://browser.primatelabs.com/geekbench3/compare/1014854?baseline=1017147

 

[MisterLilBig]

Here is a real comparison against a real AArch32 build and CPU (on Cortex-A57): http://browser.primatelabs.com/geekbench3/compare/1014854?baseline=1017147

False. That's using the "LITTLE" 4 cores Cortex-A53 at 1.3Ghz...


And if you think its using all 8 cores, do you think GeekBench has gotten GTS support? Otherwise, that's the score of Cortex-A53 at 1.3GHz.

EDIT: Sad part is, look at that much higher multicore score of the A53!!
EDIT2: Starting to think NV messed up big time.

 

[Nothingness]

False. That's using the "LITTLE" 4 cores Cortex-A53 at 1.3Ghz...


And if you think its using all 8 cores, do you think GeekBench has gotten GTS support? Otherwise, that's the score of Cortex-A53 at 1.3GHz.

Check the processor Id. And if you think A53 can reach that performance get ready to be very disappointed.

 

[MisterLilBig]

Check the processor Id. And if you think A53 can reach that performance get ready to be very disappointed.

Erm, I did. What am I suppose to find?

 

[Nothingness]

Erm, I did. What am I suppose to find?

3335=0xd07 aka ID of A57.
http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.ddi0488g/way1382037523978.html

BTW there is a slide from ARM comparing GB perf of A53 vs A15 vsA57: http://www.anandtech.com/show/7995/arm-shares-updated-cortex-a53a57-performance-expectations

Now cross check against A15 GB results on the site

 

[MisterLilBig]

Alright, comparing with an HTC Desire 510(and x2, its a 4 core A53), its similar. But still looks like its running at 1.3GHz tho. By your link, the A57 should be quite more performant than the A53, in 32bit mode. Yet, the 5433 is not that much higher than a HTC Desire 510 x2.

And that score is still a Denver running at 2.5GHz, which is not on the Nexus 9.

Edit: And that's a phone vs tablet comparison.
Edit2: Also, considering the scaling and number of cores. Isn't 4k multithread score on the Exynos 5433 rather...half...of what 8 cores should put out? Seems rather obvious.

 

[ams23]

There are now many Geekbench 3 AArch32 results for Nexus 9:

http://browser.primatelabs.com/geekbench3/search?q=htc+nexus+9

The single-core score is as high as ~ 2000 while the dual-core score is as high as ~ 3300-3500! That is incredibly good considering these are 32-bit results.

With respect to graphics, it appears that Google Nexus 9 is able to match the performance of Shield tablet in GFXBench 3.0 Manhattan Off-screen test, and is only 1 fps behind iPad Air 2 in this test:

https://gfxbench.com/device.jsp?benchmark=gfx30&os=Android&api=gl&D=Google Nexus 9

 

[Nothingness]

There are now many Geekbench 3 AArch32 results for Nexus 9:

http://browser.primatelabs.com/geekbench3/search?q=htc+nexus+9

The single-core score is as high as ~ 2000 while the dual-core score is as high as ~ 3300-3500! That is incredibly good considering these are 32-bit results.

If the 64-bit scaling is similar to Apple A8 then expect ~10% better in integer, ~50% in FP and ~10% in memory (cf this message).

OTOH given the way Denver works, it's quite difficult to predict anything.