The Coming War: ARM versus x86

[ajaidev]

"In this in-depth analysis, we will discuss the emerging competition between ARM and x86 microprocessors. Led by the Intel Atom, x86 chips are quickly migrating downwards into embedded, low-power environments, while ARM CPUs are beginning to flood upwards into the more sophisticated and demanding market spaces currently owned by x86 processors. The central focus of this report will be an extensive compute performance comparison between the ARM Cortex-A8 versus the new Intel Atom N450, the new VIA Nano L3050 and, for historical perspective, an old AMD Mobile Athlon based upon the Barton core. The Apple iPad A4 system-on-chip [SoC] is equipped with a 1GHz ARM Cortex-A8."

http://www.brightsideofnews.com/news....aspx?pageid=0

Freescale i.MX515 geez they could have done better....

[saaya]

the ARM Cortex-A8 sample that we tested in the form of the Freescale i.MX515 lived in an ecosystem that was not competitive with the x86 rivals in this comparison. The video subsystem is very limited. Memory support is a very slow 32-bit, DDR2-200MHz.

uhm... then why did you chose such a cr4ppy and limited platform if you know its a cr4ppy and limited platform?

[gOJDO]

The Coming Joke: something VS x86.

[TiN_]

ARM FTW in embedded, it's stupid to have x86 hungry monsters with chipsets and such stuff when less that few watts needed

[Glow9]

Why can't we go beyond x86? How come there isn't a open process? Who is supporting this besides the few companies that have rights to it?
It seems kinda limited :/

[-Sweeper_]

Why can't we go beyond x86? How come there isn't a open process? Who is supporting this besides the few companies that have rights to it?
It seems kinda limited :/

microsoft

[freeloader]

It really is time to move beyond X86 though. X86 has become bloated over the past few years.

[kl0012]

It really is time to move beyond X86 though. X86 has become bloated over the past few years.

And what exactly will replace x86? Those ARM chips can't even reach a performance of 14 years old pentium.

[Hell Hound]

128 bit,Quad channel def will not be slower.

[Russian]

uhm... then why did you chose such a cr4ppy and limited platform if you know its a cr4ppy and limited platform?

Because, nvidia currently doesn't have full linux support on the Tegra 2 testing platform to the point where we could have properly benchmarked it against everything else. We contacted nvidia, but they just weren't ready with it.

This is the reason why we want to do an updated article with the latest and greatest from everyone in the ARM and x86 camps.

[haylui]

x86 might not be ideal for embedded systems but x86 now has became a more RISC like design than ever
with the integrated GPU, like Llano, I believe this would bring x86 to a new high

[ajaidev]

Truth be told Tegra 2 seems great, dual A9's with OoO execution they have the power to turn the x86 based Atom/Neo/Nano on their heads. Just hope they tap in the 3D potential for UI and games that will be big for them....

[Behemot]

It's more interesting like an ultra-mobile CPU comparison. VIA is the best of all. I just do not understand why it has so low penetration. Oh, maybe Intel's practices.

Anyway. HP and Dell said they piss on smaller than 10'' devices, so we should see more AMD and VIA based systems. Definitelly good for consumer. Some people claim 8''-10'' device is enough for them, but I am sure they will take 11-12'' device if it was for the same price as it will have bigger RAM, HDD, battery, maybe even place for optical drive etc. These just do not know it is better until they try...and then they won't go back

[Helloworld_98]

And what exactly will replace x86? Those ARM chips can't even reach a performance of 14 years old pentium.

last time I checked 1.5GHz snapdragon was knocking on the door of the atom N270, in a way smaller chip.

size it up to as big as an i7 and you've got a lot of cores and GHz to work with

[Power5]

I am sorry but how the hell do I read that graph on the first page? That is the most illegible nonsense I have seen.

VIA dominates just about any bench. How can that be possible? Are these benches picked perfectly for VIA or is it just that good? Need to get that arch on 45nm process ASAP.

[Hans de Vries]

And what exactly will replace x86? Those ARM chips can't even reach a performance of 14 years old pentium.

Hmmmm, for a good old, cache based, benchmark: (performance/core)

                       MHz  VAX mips

1996 Pentium Pro       150     235
2010 Arm Cortex A9    2800    5600
2008 Celeron C2 M     2000    5275
2008 Core 2 Duo       2400    6446    
2009 Phenom II        3000    7615    
2009 Core i7          3460    9978

A single Arm Cortex A9 core* on GF's 28nm process is TEN times smaller as a
single 32 nm Westmere core showing how bloated x86 cores have become...

Regards, Hans

(* Including the 128 bit NEON SIMD unit)

[Behemot]

The truth is we see whole bunch of synthetics benchmarks, most of them only measuring some kind of throughput. We need real applications, like Office or whatever startup, some small music/video encoding, playing video, file compression etc. Just what people do every day...

[Hans de Vries]

The truth is we see whole bunch of synthetics benchmarks, most of them only measuring some kind of throughput. We need real applications, like Office or whatever startup, some small music/video encoding, playing video, file compression etc. Just what people do every day...

Yes, but such benchmarks depend on both the processor core AND the
entire memory hierarchy (caches, main memory, bandwith, pre-fetchers).

In case of the ARM core there is a different memory hierarchy for each
customer implementation.

Regards, Hans

[kl0012]

Hmmmm, for a good old, cache based, benchmark: (performance/core)

                       MHz  VAX mips

1996 Pentium Pro       150     235
2010 Arm Cortex A9    2800    5600
2008 Celeron C2 M     2000    5275
2008 Core 2 Duo       2400    6446    
2009 Phenom II        3000    7615    
2009 Core i7          3460    9978

A single Arm Cortex A9 core* on GF's 28nm process is TEN times smaller as a
single 32 nm Westmere core showing how bloated x86 cores have become...

Regards, Hans

You're comparing non existing processor in not relevant benchmark... (Not to mention that 2 DMIPS/Mhz is just a pick theoretical rate of instruction per cycle in Cortex-A9). Well, good luck and keep dreaming.

(* Including the 128 bit NEON SIMD unit)

Which is only 1 instruction per cycle capable... with unknown latency/troughput without support for DP arithmetics in current impementation.
Sandy Bridge two 256 bit AVX instructions per cycle any one...?

Yes, but such benchmarks depend on both the processor core AND the
entire memory hierarchy (caches, main memory, bandwith, pre-fetchers).

How is this "bench"? This is a theoretical rate of instruction per cycle.

[BrowncoatGR]

i think the limitations of the used ARM platform make the whole article rather pointless. Would have been nice to see a comparison with netbook class ARM platforms like Snapdragon or Tegra

[Russian]

i think the limitations of the used ARM platform make the whole article rather pointless. Would have been nice to see a comparison with netbook class ARM platforms like Snapdragon or Tegra

Trust me, we tried. When a company isn't responsive to your requests for something like that initially you just go on without them. Our goal was to show what we want to show. We really want the chance to run Snapdragon and Tegra, but as it stands right now they're not really ready yet to work with us on it.

[Manicdan]

thanks for trying, would be nice to redo this type of article in 2 years or so when both sides have 28nm being used in products

[Hans de Vries]

You're comparing non existing processor in not relevant benchmark... (Not to mention that 2 DMIPS/Mhz is just a pick theoretical rate of instruction per cycle in Cortex-A9). Well, good luck and keep dreaming.

Yes, you were off by a factor 20+ or so and now you're complaining
about a "non-existing processor" (announced quite a while ago and
discussed recently at the same website as the OP post).

http://www.brightsideofnews.com/news...a9-design.aspx

Do you realize that the whole ARM notebook frenzy started with the
prospect of the quad core ARM cortex A9 on 32/28 nm processes?
What we now see are the system level preparations for the wave
that will follow. Good luck and keep denying.

Which is only 1 instruction per cycle capable... with unknown latency/troughput without support for DP arithmetics in current impementation.
Sandy Bridge two 256 bit AVX instructions per cycle any one...?

Actually the future product Sandy Bridge does two 128 bit instructions per
half cycle. It works at double the processor clock. Otherwise the Sandy
Bridge core would have grown even further as it already did.

How is this "bench"? This is a theoretical rate of instruction per cycle.

It is a "core bench" It tests the performance of the core itself in isolation
without testing the memory interface. There is nothing theoretically in here.
It sufficient that the application fits in the cache.


Regards, Hans

[Motiv]

Quick question, is ARM a sort of open-sourced processor. Well not open sourced but basically an IP anyone can purchase, improve and produce? Also those performance figures you show Hans, do you have any that include the atom as a comparison?

I'm not really clued up on ARM, although I should fly some sort of patriotic flag being british. I seem to see more and more of these chips being used in mobile devices or one sort of another, are they actually any good or are they just a sort of stop gap CPU until Intel seriously get involved?

[Hans de Vries]

I knew I did something wrong. Its 2500 DMIPS/GHz instead of 2000, so
the table now becomes (performance per core), including an Atom result.

                         MHz  VAX mips

1996 Pentium Pro         150     235
2010 Arm Cortex A9      2800    7000
2009 Atom (2 threads)   1600    4035
2008 Celeron C2 M       2000    5275
2008 Core 2 Duo         2400    6446    
2009 Phenom II          3000    7615    
2009 Core i7            3460    9978

Regards, Hans

http://www.arm.com/products/processo.../cortex-a9.php
http://www.ocworkbench.com/2009/zota...-review/b1.htm
http://www.ocworkbench.com/2009/zota...-review/b1.htm