11 replies to this topic

[SpleenBeGone]

So, having all these new things around me, and being a nerd, I feel the need to benchmark them. I'm currently using an android app called AnTuTu, but if there's one that's cross platform, someone please let me know. Here's my scores so far, feel free to post your own.

Nexus 7 32gb:
Total: 10300
Ram: 1893
CPU Integer: 3507
CPU Float-point: 2525
2D Graphics: 296
3D Graphics: 1233
Database IO: 535
SD Card Write: (11.7MB/s) 117
SD Card Read: (31.1 MB/s) 194
CPU Frequency: 1300Mhz

LG Optimus G

Total: 14086
Ram: 3334
CPU Integer: 4764
CPU Float-point: 3548
2D Graphics: 294
3D Graphics: 1249
Database IO: 555
SD Card Write: (29.3MB/s) 150
SD Card Read: (>50 MB/s) 197
CPU Frequency: 1512Mhz

I know the Optimus should run at 1.5Ghz, not sure why it's clocked down there.

Edit: Clocked down due to heat apparently, I amended the score with it at the proper clock rate.

[idk]

inb4 laggy

[Andux]

Some time ago I posted a reply to this (twice) but it did not appear. This might be because I include links to my numerous benchmarks and results. Following is the text, without the links. All can be found on Googling (like just roylongbottom or with a name like Whetstone).

 

You might like my FREE, too many for non-nerds, benchmarks covering Windows, Linux and Android. For Android here are 14 single core benchmarks and 8 using multithreading. See following for details, downloads and results: Results are in MFLOPS, MIPS and MB/second, not meaningless numbers. Most use Native Code with a couple Java versions for comparison. Some are supplied in different flavours - Linpack via Java, native DP, native SP and native old HW.

 

Links to Android benchmarks

 

First set are the Classic Benchmarks that were the first programs that set standards of performance for computers. They are (original standard) Whetstone (Floating point benchmark for minicomputers), Dhrystone (Integer benchmark for UNIX systems), Livermore Loops (Numeric benchmark for supercomputers) and Linpack (Floating point benchmark for workstations). Then there are some that measure bus, cache and memory speeds.

 

These and others are available to run on PCs under Windows and Linux from:

 

Links to Linux benchmarks

 

A summary of all Classic Benchmark results are in:

 

Links to benchmark results (including cross platform)

 

and for old minicomputers and mainframes (Whetstone Benchmark History and Results

): links

 

 

Roy

[Andux]

Cross Platform Benchmarks

 

The first cross platform program worthy of consideration is the Whetstone Benchmark. This appears to have been used by ARM in the 1990’s and later.

 

Whetstone Benchmark

 

The original, written in Fortran, was the first general purpose benchmark that set industry standards of computer performance. It was released in 1972, based on research by Brian Wichmann, and produced by Harold Curnow. Later updates became my responsibility. The three of us were UK Government employees. Speed was measured in terms of Million Whetstone Instructions Per Second (MWIPS). Later, in order to identify compiler over-optimisation, I introduced speeds of individual tests, shown as MOPS or MFLOPS - Millions of Operations or Floating Point Operations Per Second. My results for all sorts of programming languages, mainly via Windows, Linux and Android are in the following.

 

http://www.roylongbo...one results.htm 

 

Then Whetstone History and Results (mainframes, minicomputers, supercomputers 1970s to 1990s in:

 

http://www.roylongbo...k/whetstone.htm

 

I have produced two versions for Android. The first is an all Java variety (Java Whetstone.apk) and the second one has a Java front end with the benchmark using compiled C?C++ code (NativeWhetstone.apk). These can be downloaded from and need a Settings, Security optin to allow installation of non-Market applications.

 

http://www.roylongbo... benchmarks.htm

 

Following are example results where #A is a Nexus 7, #B a Galaxy SIII (new version A9), QU-S4  Qualcomm Snapdragon S4 (some performance differences to ARM) with Atom and Core 2 PCs via Linux. Harmonic mean results are shown for MFLOPS, COS/EXP MOPS and Integer tests. Note: some CPUs do not appear to run at the specified MHz and other need a Setting to force maximum speed.

 

                                       Func     INT     MWIPS
  CPU            MHz   MWIPS  MFLOPS    MOPS    MOPS   per MHz
         JAVA
  926EJ          800      14       3     0.3       5      0.02
  v7-A8         1000     138      22     3.3      51      0.14
  v7-A9          800     224      44     6.8      45      0.28
  v7-A9 #A      1300     348      73     8.4      81      0.27
  v7-A9 #B      1400     400      84     9.8      93      0.29

  QU-S4         1500     487      81    13.5     127      0.32

  Atom          1666     621     248    12.0     183      0.37
  Core2         2400    1925     635    34.8     739      0.80

         NATIVE
  926EJ          800      31      11     0.4      88      0.04
  v7-A8         1000     288      89     4.4     374      0.29
  v7-A9          800     687     156    11.7     814      0.86
  v7-A9 #A      1300    1115     259    18.6    1345      0.86
  v7-A9 #B      1400    1334     301    24.9    1569      0.95

  QU-S4         1500    1040     288    16.0    1590      0.69

  Atom          1666     769     313    10.1    1080      0.46
  Core 2        2400    2560     753    40.3    2181      1.07

 

 

At least on a per MHz basis, the latest ARM CPUs can be as fast as Intel processors. 

 

 Roy

 

 

[SpleenBeGone]

I have a Snapdragon S4 in my phone, and I can confirm it's a powerhouse. 

 

I find your last statement a bit misleading though, as you only go up to a Core2 processor and don't list what architecture it is. Conroe<Kentsfield<Yorkfield for example. You'll also find that clock for clock, the Sandy Bridge and Ivy Bridge pretty well blow Core2 out of the water, and Haswell coming up is expected to be a pretty nice increase as well. 

[Andux]

> as you only go up to a Core2 processor and don't list what architecture it is. Conroe<Kentsfield<Yorkfield for example. You'll also find that clock for clock, the Sandy Bridge and Ivy Bridge pretty well blow Core2 out of the water<

The Whetstone benchmark does not depend on L2/L3 cache size and performance appears to be consistent across Core 2 and i7, at least on the Windows version shown in the following results. Other benchmarks vary according to CPUID. In this case Sandy Bridge seems to have sunk a bit. These results are from CPUs running at the Turbo Boost frequency.

 

                                              Func     INT     MWIPS
   CPU                 MHz   MWIPS  MFLOPS    MOPS    MOPS   per MHz  CPUID

 

  Core 2 Duo          2400    2057     499    42.8     685      0.86  10676
  Core i7 930         3066    2496     576    54.5     830      0.81  106A5
  Core i7 860         3466    2790     653    61.5     910      0.80  106E5
  Core i7 3930K       3800    3004     707    64.9    1122      0.79  206D7

 

 

  Roy

 

[SpleenBeGone]

Well that's rather interesting. Do you have any info on Ivy Bridge stuff? I imagine the trigate transistors and die shrink would make a difference. 

[Andux]

I don’t have any Ivy Bridge results yet, only up to Sandy Bridge. For the latter see the following overnerdified results  with %MIPS/MHz and %MFLOPS/MHz for CPU/L1. L2, L3 and RAM data for most of my single CPU benchmarks. The Core i5 is Sandy Bridge, not much different on these ratios.

 

At the start are CPUIDs and other things. Note differences on cache sizes, Bus GT/s and Mem GB/s, including some Ivy Bridge. Also note Celerons and Zeons with same CPUID as Core systems.

 

http://www.roylongbo...uk/cpuspeed.htm 

 

Roy

[Andux]

The second cross platform test is the Dhrystone Benchmark. This is a C integer only program and has been around since 1984. Speed was originally measured in Dhrystones per second. This was later changed to VAX MIPS by dividing Dhrystones per second by 1757, the DEC VAX 11/780 result, the latter being regarded as the first 1 MIPS minicomputer. This is perhaps ARM’s favourite benchmark, where they quote 2.5 MIPS/MHz for Cortex-A9. This is much better than with my benchmark but manufacturer’s compilers are inevitably tuned to over-optimise.

My benchmarks are Dhrystone2.apk and Dhry2Nopt.apk, downloadable from

http://www.roylongbo... benchmarks.htm

Result on PCs can be found in:

http://www.roylongbo...one results.htm

The original Dhrystone 1 and 2 results are in

http://www.dunningto...lic/dhrystone.c
and
http://performance.n....data.col0.html

The latter includes 1996 ARM Risc CPU result, included in the following table. Results for PCs are also given, to show variation of using different compilers.

Dhrystone 2 Optimised

 

 

 

  PC Windows    MHz   VAX   MIPS     Android Device      MHz   VAX    MIPS
     /Linux          MIPS   /MHz                              MIPS    /MHz

 

  Atom            1666    1828     1.1    926EJ Tablet             800     356    0.45

  Core i5 2467M   2300    4752     2.1    v7-A9 V0x1 Tablet ##     800     962    1.20
  Core 2          2400    6446     2.7    v7-A9 V0x2 Galaxy SII   1200    1491    1.24
  Core i7 930     3066    8684     2.8    v7-A9 V0x3 Tablet ##    1500    1650    1.10
  Core i7 860     3460    9978     2.9    v7-A9 V 0x2 Nexus 7     1300    1610    1.24
  Core i7 3930K   3800   11197     2.9    v7-A9 V 0x3 Galaxy SIII 1400    1937    1.38
  Later
  Core 2 32 bit   2400    8094     3.4    Old 1996 Acorn Risc
  Core 2 64 bit   2400   12600     5.3
  Linux                                   ARM SA110 Acorn PC       200     204    1.02
  Atom 32 bit     1666    2055     1.2    ARM SA110 Acorn PC        30    18.4    0.61
  Atom 64 bit     1666    2704     1.6    ARM 3 Acorn 5000          33    13.8    0.42
  Core 2 32 bit   2400    5852     2.4
  Core 2 64 bit   2400   12265     5.1    ## not running at specified  MHz

 

                     WANTED ANDROID RESULTS FOR OTHER CPUs

 

 

[Andux]

Linpack Benchmark

 

The Linpack Benchmark was produced from the "LINPACK" package of linear algebra routines. It became the primary benchmark for scientific applications, particularly under Unix, from the mid 1980's, with a slant towards supercomputer performance. The original double precision C version, used here, operates on 100x100 matrices. Performance is governed by an inner loop in function daxpy() with a linked triad dy[i] = dy[i] + da * dx[i], and is measured in Millions of Floating Point Operations Per Second (MFLOPS).

 

This benchmark, from GreenComputing” seems to be the most popular one for Android from Google Play. However, it is based on Java calculations and is extremely slow compared with other systems. Results for the original N=100 and others for MP systems can be found in: 

 

ftp://ftp.idsa.prd.fr/pub/netlib/benchmark/performance.pdf

 

and my results on PCs via Windows and Linux, plus Android devices are in:
 
http://www.roylongbo...ack results.htm

 

I have five versions available that use one CPU core and are downloadable via:

 

http://www.roylongbo... benchmarks.htm

 

LinpackJava.apk   - all Java
Linpackv5.apk      - compiled native code for old floating point instructions
Linpackv7.apk      - compiled native code using later vfpv3 instructions
LinpackSP.apk      - as v7 but using faster single precision floating point
NEON-Linpack.apk - using NEON SIMD vector instructions (single precision)

 

Example results are below. The benchmark uses L2 cache sized data, where the HTC One X with a Qualcomm Snapdragon S4 is faster than the ARM-A9 CPUs. Unlike Intel, single precision calculations are faster than double precision on the ARM CPUs but the main line Core processors are far superior.

 

 

 Linpack MFLOPS
                                                                MFLOPS/MHz
 Device            ARM      MHz     V7     SP   NEON   Java      V7   SIMD

 

 Tablet          926EJ      800      6     10    N/A      2    0.01
 Huawei u8800    v7-A8      800     80                         0.10
 Tablet          v7-A9      800    101    129    256     33    0.13   0.32
 HTC One X       v7-A9     1500    171                         0.11
 Tablet          v7-A9     1500    156    205    382     57    0.10   0.25
 Asus TF700      v7-A9     1600    196                         0.12
 Nexus 7         v7-A9     1300    151    201    376     56    0.12   0.29
 HTC One X       QU-S4     1500    255                         0.17
 Galaxy SIII     v7-A9     1400    184    236    454     57    0.13   0.32

 

 Linux                             Opt
 Atom                      1666    204    216           118    0.12
 Core 2                    2400   1288                  901    0.54

 

 Windows                           Opt
 Atom                      1666    183                  119    0.11
 Core 2                    2400   1315                  551    0.55
 Core i7 930     Turbo     3066   1765                         0.58
 Core i7 860     Turbo     3460   2004                         0.58
 Core i7 3930K   Turbo     3800   2530                         0.67

                                  SSE2
 Core 2 64 bit             2400   1602                                0.67
 Core i7 3930K   O'Clkd    4720   3928                                0.83

 

 Roy

[Andux]

How much faster is your phone  than the $7M Cray 1 Supercomputer?

 

Livermore Loops

 

This original main benchmark for supercomputers was first introduced in 1970, initially comprising 14 kernels of numerical application, written in Fortran. This was increased to 24 kernels in the 1980s. Performance measurements are in terms of Millions of Floating Point Operations Per Second or MFLOPS. The kernels are executed three times with different double precision data array sizes. Following are overall MFLOPS results for various systems, Geometric Mean being the official average performance. [Reference - F.H. McMahon, The Livermore Fortran Kernels: A Computer Test Of The Numerical Performance Range, Lawrence Livermore National Laboratory, Livermore, California, UCRL-53745, December 1986]

                   ---------------- MFLOPS ---------------              
 CPU          MHz  Maximum Average Geomean Harmean Minimum  Date

 CDC 7600      36.4   7.3     4.2     3.9     2.5     1.4   1974 * 
 Cray 1A       80    83.5    25.8    14.4     7.9     2.7   1980 * 
 Cray 1S       80    82.1    22.2    11.9     6.5     1.0   1985    
 CDC Cyber 205 50   146.9    36.4    14.6     5.0     0.6   1982 * 
 Cray 2       244   146.4    36.7    14.2     5.8     1.7   1985    
 Cray XMP1    105   187.8    61.3    31.5    15.6     3.6   1986    

 

             * Fewer than 24 Kernels, Date is year measured   
 
The  Android benchmark is LivermoreLoops.apk, available from:

 

http://www.roylongbo... benchmarks.htm

 

Android results below show Geometric Mean speed up to 15 times Cray 1, but not as effective as desktop PCs. These kernels use L1 and L2 caches, penalising the Snapdragon with the smaller L1 cache (indicated on memory benchmarks).

 

                              ------------- MFLOPS ------------
 Device        CPU   MHz     Max Average Geomean Harmean     Min

 

 Tablet      926EJ   800      10       6       5       5       2
 Galaxy S    v7-A8  1000      61      37      35      32      13
 Tablet      v7-A9   800     253     129     115     102      47
 Nexus 7     v7-A9  1300     392     202     181     161      68
 Tablet      v7-A9  1500     397     208     186     165      75
 HTC One S   QU-S4  1500     418     228     194     161      65
 Galaxy SIII v7-A9  1400     456     247     221     196      85

 Linux
 Netbook     Atom   1666     465     212     185     157      50
 Desktop    Core 2  2400    2385    1038     806     582     161
 Desktop    i7 930  3066    4336    1574    1317    1105     410

 

                 QU-S4 = Qualcomm Snapdragon S4

 

 Roy

[Andux]

Memory Benchmarks

 

 The most noticeable performance differences between the various implementations of ARM processor cores are revealed via my memory benchmarks. The three of them all measure MegaBytes per second transferring from caches and RAM, with data sizes 16. 32. 64. 126. 256. and 512 KB, then 1, 4, 16 and 64 MB.

 

MemSpeed - Calculations are x[m]=x[m]+s*y[m] and x[m]=x[m]+y[m], using double and single precision floating point and x[m]=x[m]+s+y[m] and x[m]=x[m]+y[m] with integers.

 

BusSpeed - This benchmark reads data using AND instructions and is designed to identify transferring data in bursts over buses. The program starts by reading a word (4 bytes) with an address increment of 32 words (128 bytes) before reading another word. The increment is reduced by half on successive tests, until all data is read. Different burst sizes can apply to transfers from RAM and burst might or might not apply via caches.

 

RandMem - carries out four tests with serial and random read and read/write tests, using 32 bit integers. Serial and random address selections use the same complex indexing structure. Random access speeds are influenced by burst reading.

 

Further details, full results and download links are provided in:

 

http://www.roylongbo... benchmarks.htm   

 

As with other benchmarks, comparison of these ones can show that CPUs don’t always run at the specified MHz. They also differences between earlier and later Cortex-A9 processors and those using Qualcomm Snapdragon S4s.

 

The Nexus 7 and Galaxy SIII both have quad core Cortex-A9 CPUs, the SIII having a later version. There are claims that this has 128-bit internal buses, instead of 64-bit and is probably why speeds via L2 cache are up to 25% faster. The SIII also have dual channel RAM with serial data transfer speeds three times faster.

 

Comparing a HTC One X Snapdragon S4 with the SIII identifies significant differences. The former has a smaller L1 cache then, adjusting for a constant MHz, is slower than the Cortex-A9 on integer calculations and much slower on all sequential data transfers form RAM (not dual channel?). On the other hand, the S4 is up to twice as fast on floating point calculations with data from caches and can be faster on random/skipped sequential data transfers from RAM, due to different burst reading characteristics. Back to disadvantages, the S4 can suffer from severe degradation on reading random/skipped sequential data from caches due to apparent burst reading.

 

We could say that these benchmarks show that we can’t simply say that one phone or tablet is faster than another with a variation of technology. In this case, a SIII can be between half speed to 2.5 times faster than the S4.

 

Roy