Intel Core i5 750, Core i7 870 Linux Benchmarks

[justapost]

With my results using open64 on a i7-920 without overclocking(running ubuntu):

$ head -8 /proc/cpuinfo
processor : 0
vendor_id : GenuineIntel
cpu family : 6
model : 26
model name : Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz
stepping : 4
cpu MHz : 2672.704
cache size : 8192 KB

$ ./stream-open64
-------------------------------------------------------------
STREAM version $Revision: 5.9 $
-------------------------------------------------------------
This system uses 8 bytes per DOUBLE PRECISION word.
-------------------------------------------------------------
Array size = 20000000, Offset = 0
Total memory required = 457.8 MB.
...
Function Rate (MB/s) Avg time Min time Max time
Copy: 22410.3334 0.0143 0.0143 0.0143
Scale: 22282.7187 0.0144 0.0144 0.0144
Add: 22511.9469 0.0230 0.0213 0.0234
Triad: 20943.1595 0.0233 0.0229 0.0234

Those are triple channel results right? Max possible with dual channel ddr3 1333 would be 21200MB/s. I'd be interested in dual channel results, I expect something around 17GB/s. On an amd setup you must run the nb at around 2.6GHz and the cpu abit faster 2.8GHz to hit the max which was around 15.7GB/s here with DDR3 1333 CL7 (stream build with icc 11.1 and openmp support).

Update: Looking at your results it seemed to me open64 compilers generate as efficient code as icc 11.1 so I grabed the actual build and ran a comparison here.

PII 955BE 3.2GHz NB 2GHz MEM 2xDDR1333 Unganged CL7

ICC 11.1

Function      Rate (MB/s)   Avg time     Min time     Max time
Copy:       13223.4215       0.0026       0.0024       0.0150
Scale:      13261.3109       0.0025       0.0024       0.0090
Add:        13726.5011       0.0036       0.0035       0.0048
Triad:      13788.5482       0.0036       0.0035       0.0050

Open64 4.2.1

Copy:        8859.2560       0.0036       0.0036       0.0037
Scale:       8712.0426       0.0037       0.0037       0.0037
Add:         9541.0925       0.0050       0.0050       0.0051
Triad:       9749.9439       0.0056       0.0049       0.0111

GCC-4.3.1

Copy:        8820.2489       0.0041       0.0036       0.0055
Scale:       8544.5460       0.0041       0.0037       0.0054
Add:         9597.9497       0.0053       0.0050       0.0055
Triad:       9632.8513       0.0059       0.0050       0.0100

PII 955BE 3.2GHz NB 2.6GHz MEM 2xDDR1333 Unganged CL7


ICC 11.1

Function      Rate (MB/s)   Avg time     Min time     Max time
Copy:       15116.3113       0.0023       0.0021       0.0089
Scale:      15794.0372       0.0021       0.0020       0.0029
Add:        15185.2913       0.0032       0.0032       0.0035
Triad:      15320.4925       0.0032       0.0031       0.0050

Open64 4.2.1

Function      Rate (MB/s)   Avg time     Min time     Max time
Copy:        9624.1021       0.0054       0.0033       0.0216
Scale:       9329.0977       0.0035       0.0034       0.0035
Add:        10278.5823       0.0047       0.0047       0.0047
Triad:      10478.1197       0.0046       0.0046       0.0047

GCC-4.3.1

Function      Rate (MB/s)   Avg time     Min time     Max time
Copy:        9445.3011       0.0034       0.0034       0.0034
Scale:       9297.4320       0.0035       0.0034       0.0035
Add:        10225.8529       0.0047       0.0047       0.0048
Triad:      10405.5505       0.0046       0.0046       0.0047

I used openCC -fopenmp -O2 -o stream-o64 stream.c to build the open64 version. Seems I'm doing something wrong here.

[AdrenalineJunky]

The i7 750 can be faster than the i7 920 because the i5 can use 5 levels of turbo boost and the i7 only 3. Each step is 133 mhz, some are only available with 1 or 2 cores. A benchmark like povray which only runs on 1 core should show it.

except that in this test turbo boost was disabled.

[Kano]

The i7 750 can be faster than the i7 920 because the i5 can use 5 levels of turbo boost and the i7 only 3. Each step is 133 mhz, some are only available with 1 or 2 cores. A benchmark like povray which only runs on 1 core should show it.

[deanjo]

My best guess is that the compilers used under windows are more aware of the special features of the newest chips. Things like the double wide SSE, increased number of micro-ops possible, and various other optimizations. That and the combination of turning off the auto overclocking (mentioned in some detail in the p55 article) combine to show AMD with a substantial advantage.

So while this advantage is real *today*, with problems with intel's turbo boost, and compiling binaries with the default gcc.

There are compilers aware of such features. Portland Group, Pathscale, and the free open64. As an example compare the i7-920 results at:

http://www.phoronix.com/scan.php?page=article&item=intel_lynnfield&num=13

With my results using open64 on a i7-920 without overclocking(running ubuntu):

$ head -8 /proc/cpuinfo
processor : 0
vendor_id : GenuineIntel
cpu family : 6
model : 26
model name : Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz
stepping : 4
cpu MHz : 2672.704
cache size : 8192 KB

$ ./stream-open64
-------------------------------------------------------------
STREAM version $Revision: 5.9 $
-------------------------------------------------------------
This system uses 8 bytes per DOUBLE PRECISION word.
-------------------------------------------------------------
Array size = 20000000, Offset = 0
Total memory required = 457.8 MB.
...
Function Rate (MB/s) Avg time Min time Max time
Copy: 22410.3334 0.0143 0.0143 0.0143
Scale: 22282.7187 0.0144 0.0144 0.0144
Add: 22511.9469 0.0230 0.0213 0.0234
Triad: 20943.1595 0.0233 0.0229 0.0234

For comparison a Phenom II X4 810 (2.6 GHz) + DDR3-1333 using the same binary:
Function Rate (MB/s) Avg time Min time Max time
Copy: 12455.2457 0.0258 0.0257 0.0258
Scale: 12369.4995 0.0259 0.0259 0.0260
Add: 12539.4940 0.0384 0.0383 0.0387
Triad: 12442.0063 0.0387 0.0386 0.0387

Using gcc-4.4.1 on i7-920:
$ ./stream-gcc-4.4.1
Function Rate (MB/s) Avg time Min time Max time
Copy: 14374.3618 0.0223 0.0223 0.0224
Scale: 14416.3573 0.0222 0.0222 0.0223
Add: 15624.5172 0.0308 0.0307 0.0308
Triad: 15801.4749 0.0304 0.0304 0.0

Same on X4 810:
Function Rate (MB/s) Avg time Min time Max time
Copy: 8490.9773 0.0378 0.0377 0.0380
Scale: 8485.1263 0.0379 0.0377 0.0383
Add: 9569.1637 0.0503 0.0502 0.0508
Triad: 9573.1679 0.0505 0.0501 0.0528

For fun I'll run just 3 copies on the X4-810 to simulate using an X3:
Function Rate (MB/s) Avg time Min time Max time
Copy: 8326.3684 0.0386 0.0384 0.0389
Scale: 8329.6239 0.0386 0.0384 0.0389
Add: 9358.3690 0.0514 0.0513 0.0517
Triad: 9346.5083 0.0514 0.0514 0.0518

So as you can see using current compilers (default under windows or optionally under linux) can yield large differences in performance, in this case over 2x for the i7-920.

Should be noted that Pathscale maybe no longer available in the near future. It all depends what happens now that Cray has bought it.

[BillBroadley]

AMD vs Intel on Linux vs Windows

when i saw they were comparing an X3 i rolled my eyes...

when i saw the results i nearly fell out of my chair in amazement.

WHAT? how does that even happen? in windows the i5 was obliterating the phenom 965....

My best guess is that the compilers used under windows are more aware of the special features of the newest chips. Things like the double wide SSE, increased number of micro-ops possible, and various other optimizations. That and the combination of turning off the auto overclocking (mentioned in some detail in the p55 article) combine to show AMD with a substantial advantage.

So while this advantage is real *today*, with problems with intel's turbo boost, and compiling binaries with the default gcc.

There are compilers aware of such features. Portland Group, Pathscale, and the free open64. As an example compare the i7-920 results at:

http://www.phoronix.com/scan.php?page=article&item=intel_lynnfield&num=13

With my results using open64 on a i7-920 without overclocking(running ubuntu):

$ head -8 /proc/cpuinfo
processor : 0
vendor_id : GenuineIntel
cpu family : 6
model : 26
model name : Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz
stepping : 4
cpu MHz : 2672.704
cache size : 8192 KB

$ ./stream-open64
-------------------------------------------------------------
STREAM version $Revision: 5.9 $
-------------------------------------------------------------
This system uses 8 bytes per DOUBLE PRECISION word.
-------------------------------------------------------------
Array size = 20000000, Offset = 0
Total memory required = 457.8 MB.
...
Function Rate (MB/s) Avg time Min time Max time
Copy: 22410.3334 0.0143 0.0143 0.0143
Scale: 22282.7187 0.0144 0.0144 0.0144
Add: 22511.9469 0.0230 0.0213 0.0234
Triad: 20943.1595 0.0233 0.0229 0.0234

For comparison a Phenom II X4 810 (2.6 GHz) + DDR3-1333 using the same binary:
Function Rate (MB/s) Avg time Min time Max time
Copy: 12455.2457 0.0258 0.0257 0.0258
Scale: 12369.4995 0.0259 0.0259 0.0260
Add: 12539.4940 0.0384 0.0383 0.0387
Triad: 12442.0063 0.0387 0.0386 0.0387

Using gcc-4.4.1 on i7-920:
$ ./stream-gcc-4.4.1
Function Rate (MB/s) Avg time Min time Max time
Copy: 14374.3618 0.0223 0.0223 0.0224
Scale: 14416.3573 0.0222 0.0222 0.0223
Add: 15624.5172 0.0308 0.0307 0.0308
Triad: 15801.4749 0.0304 0.0304 0.0

Same on X4 810:
Function Rate (MB/s) Avg time Min time Max time
Copy: 8490.9773 0.0378 0.0377 0.0380
Scale: 8485.1263 0.0379 0.0377 0.0383
Add: 9569.1637 0.0503 0.0502 0.0508
Triad: 9573.1679 0.0505 0.0501 0.0528

For fun I'll run just 3 copies on the X4-810 to simulate using an X3:
Function Rate (MB/s) Avg time Min time Max time
Copy: 8326.3684 0.0386 0.0384 0.0389
Scale: 8329.6239 0.0386 0.0384 0.0389
Add: 9358.3690 0.0514 0.0513 0.0517
Triad: 9346.5083 0.0514 0.0514 0.0518

So as you can see using current compilers (default under windows or optionally under linux) can yield large differences in performance, in this case over 2x for the i7-920.

[AdrenalineJunky]

when i saw they were comparing an X3 i rolled my eyes...

when i saw the results i nearly fell out of my chair in amazement.

WHAT? how does that even happen? in windows the i5 was obliterating the phenom 965....

[kotsergy]

One thing about comparison tables, it would've made more sense to arrange CPU results in order of increasing price from left to right for example, so that it is more practical and comfortable to compare the results. X3 then I7 then I5 then I7 does not make much sense.

[Apopas]

It's worth nothing that Microsoft licensed the intel compiler technology to enable more aggressive optimizations.

To achieve similar performance numbers under linux you need to run Pathscale, Portland group, or the free open64 compilers to get similar numbers. For instance using said compilers gets over double the performance with stream when compared to the phoronix posted numbers.

Is possible then, the Intel's proccessors to run better the programs that have been compiled with their own compiler?

[BillBroadley]

Windows compiler

That would have to be the MSVS series one would have to assume.

It's worth nothing that Microsoft licensed the intel compiler technology to enable more aggressive optimizations.

To achieve similar performance numbers under linux you need to run Pathscale, Portland group, or the free open64 compilers to get similar numbers. For instance using said compilers gets over double the performance with stream when compared to the phoronix posted numbers.

[sabby]

Just to add my 2 cents, I had mix feeling to see most review site post result with Turbo boost enabled. I personally would prefer to also see the result without. Nonetheless, I am quite surprised at the result posted on phoronix. Definitely will postpone buying one until at least things are sorted out, already do enough beta testing as is. Hopefully won't see a repeat of the problems with poulsbo...