How fast are different ARM CPUs?

[pierro78]

Sheeva is cortex A8 : http://extranet.marvell.com/technolo...pu_history.jsp

[u2maemo]

Originally Posted by shadowjk

Sheevaplug 1.2 GHz

Processor : Feroceon 88FR131 rev 1 (v5l)

v5l means ARMv5

[craftyguy]

Originally Posted by u2maemo

I just wonder why
arm v5(Sheevaplug 1.2 GHz)
fast than
both arm v7(omap 3420 cortex A8 600 MHz)
and arm v6(omap 2420 arm11 400 MHz)

frequency?
or
pi is not suit for performance test

My guess is that pi is more sensitive to frequency than any microarchitecture differences between ARM 5 & 8..

[shadowjk]

The description for pi says it's limited only by your machine's ram. The SheevaPlug has half a gigabyte of fast ram.

[mrebanza]

is it possible to UPGRADE any of the hardware on the N900???


RAM - PROCESSOR - ECT ? ? ?

[Rob1n]

Originally Posted by mrebanza

is it possible to UPGRADE any of the hardware on the N900???


RAM - PROCESSOR - ECT ? ? ?

No. They're all soldered onto the boards, and there's all sorts of power & thermal issues involved.

[shadowjk]

These results aren't making much sense..

SmartQ7, 667MHz s3c6410 ARMv6, original firmware 5.0

1m35.872s

[maxximuscool]

here is my N900:

~ $ cat /proc/cpuinfo
Processor : ARMv7 Processor rev 3 (v7l)
BogoMIPS : 249.96
Features : swp half thumb fastmult vfp edsp neon vfpv3
CPU implementer : 0x41
CPU architecture: 7
CPU variant : 0x1
CPU part : 0xc08
CPU revision : 3

Hardware : Nokia RX-51 board
Revision : 2101
Serial : 0000000000000000
~ $

[shadowjk]

EDIT: Added intel Atom 330 and Intel Core 2 Quad 9550 CPUs.


Another benchmark, distributed.net OGR and RC5. Clients available for arm-eabi at distributed.net, ./dnetc -bench, take fastest result.

The results for the multicore hyperthreaded Atom330 and multicore Q9550 are for a single thread on a single core.

I hope the formatting remains sane. I found no proper way of making tables on the forum...

RC5-72

Device,   Clockspeed  result,       relative to N800, per clock
N800:     400MHz    448,553 keys/sec  100%  1121 keys/s/MHz
N900:     600MHz  1,001,396 keys/sec  223%  1669 keys/s/MHz
SmartQ7:  667MHz    751,596 keys/sec  167%  1127 keys/s/MHz
Sheeva:  1200MHz  1,314,021 keys/sec  293%  1095 keys/s/MHz
Atom330: 1600MHz  3,208,177 keys/sec  715%  2005 keys/s/MHz
Q9550(1):2830MHz  9,479,121 keys/sec 2113%  3350 keys/s/MHz
Q9550(2):2830MHz 10,562,046 keys/sec 2355%  3732 keys/s/MHz

Atom330 in 32bit mode.
(1) 64bit client with less optimizations
(2) 32bit client, seems to have more optimizations

Interesting things:

N900 has 1.5 times the clockspeed of N800, yet it's over 2 times faster.
Sheevaplug has DOUBLE the clockspeed of N900, yet it's only 1.3 times as fast as N900.

OGR-NG:

Device,   Clockspeed, result,        relative to N800, nodes per clock
N800:     400MHz  1,855,943 nodes/sec  100%  4640 nodes/s/MHz
N900:     600MHz  3,642,503 nodes/sec  196%  6071 nodes/s/MHz
SmartQ7:  667MHz  3,070,970 nodes/sec  165%  4604 nodes/s/MHz
Sheeva:  1200MHz  4,986,341 nodes/sec  268%  4155 nodes/s/MHz
Atom330: 1600MHz  9,893,719 nodes/sec  533%  6184 nodes/s/MHz
Q9550(1):2830MHz 37,681,342 nodes/sec 2030% 13314 nodes/s/MHz
Q9550(2):2830MHz 46,671,019 nodes/sec 2515% 16492 nodes/s/MHz

Atom330 in 32 bit mode.
(1) 64 bit client
(2) 32 bit client with SSE2 optimizations

The overall picture is the same, the N900's CPU does more work per clcok than all the other ARM cores. Interestingly, it's not far behind Atom here. Remember, that for decent battery life in a portable, the atoms usually run at 600-800MHz.. That puts them in an interesting position when compared to Omap3...

Core2 is brutal, but we knew that. And it gobbles up insane amounts of power to keep its 1 kilogram heatsink hot...

[TA-t3]

>N900 has 1.5 times the clockspeed of N800, yet it's over 2 times faster.

That makes sense, it's OMAP3 vs. OMAP2. The only surprise is that it isn't even faster. Was the N900 version compiled with options for the architecture?

Is there floating point involved? Double- or single-precision, if so?