Pandora Cpu Question

[es87]

The Pandora's ARM CPU is 500-900 MHz. In terms of raw power, how fast would an equivalent x86 CPU be?

 

[MagicPants]

The Pandora's ARM CPU is 500-900 MHz. In terms of raw power, how fast would an equivalent x86 CPU be?

In terms of power most desktop x86s draw around 30 to 60 watts, an ARM is more in the 1 to 4 watt range.

 

[es87]

In terms of power most desktop x86s draw around 30 to 60 watts, an ARM is more in the 1 to 4 watt range.

I meant the equivalent speed (For example, the answer could be 800MHz - 1600MHz)

 

[hch]

The Pandora's ARM CPU is 500-900 MHz. In terms of raw power, how fast would an equivalent x86 CPU be?

that's not really well defined. since the A8 is a superscalar in-ordr core with SIMD extensions and 256(?)k L2 cache, i'd guess that it is roughly equivalent to an atom of the same frequency. it has the same bogomips for sure, but floating point performance depends on how many fp units it has (i think it has 2, like the atom). compared to a core-2 quad however, i'd guess it should be equivalent to a 120-200 mhz version.

then there is the memory bandwidth that is sometimes more important than raw cpu power. i thing the A8 has about 1.5GB/s, which is almost a factor 10 worse than current desktop x86 machines.

very roughly i'd expect the whole machine to perform similar to an early pentium-3 system.

 

[es87]

The Pandora's ARM CPU is 500-900 MHz. In terms of raw power, how fast would an equivalent x86 CPU be?

that's not really well defined. since the A8 is a superscalar in-ordr core with SIMD extensions and 256(?)k L2 cache, i'd guess that it is roughly equivalent to an atom of the same frequency. it has the same bogomips for sure, but floating point performance depends on how many fp units it has (i think it has 2, like the atom). compared to a core-2 quad however, i'd guess it should be equivalent to a 120-200 mhz version.

then there is the memory bandwidth that is sometimes more important than raw cpu power. i thing the A8 has about 1.5GB/s, which is almost a factor 10 worse than current desktop x86 machines.

very roughly i'd expect the whole machine to perform similar to an early pentium-3 system.


Wow, I expected better performance than a Pentium 3!

 

[fischju2000]

The Xbox has a 733mhz PIII

 

[Svartalf]

I meant the equivalent speed (For example, the answer could be 800MHz - 1600MHz)

Unfortunately it's not as simple as that. It never HAS been that simple, really. Clock speed, for example, is only relevant within a given range of a product line. (A Core Duo runs faster at lower clock speeds than a P4, for example- or more compelling, an ARM11 machine will run 1/2-1/3 as fast per clock when compared to a Cortex-A8 machine and consume the same power doing it...)

Not having done decent comparisons (Yet...) I couldn't tell you past a rough guess that it's comparable to slightly better than what Intel's doing with the Atom per clock with a fraction of the power consumption- IF you've got good efficient coding done for the CPU that doesn't trip over itself with unaligned reads/writes, etc.

 

[crasherball]

r u kidding?
we're talking about a handheld here...
pentium 3 is more then just good

 

[Exophase]

(A Core Duo runs faster at lower clock speeds than a P4, for example- or more compelling, an ARM11 machine will run 1/2-1/3 as fast per clock when compared to a Cortex-A8 machine and consume the same power doing it...)

Unless you're talking about a program that is very SIMD heavy then I have a very hard time believing this. I'd like to know how you came up with that number.

 

[Svartalf]

Wow, I expected better performance than a Pentium 3!

You're kidding, right?

Atom gets about the same as a PIII in performance. In a handheld, this is largely unheard of speed- especially at the power consumption we're getting with it. As someone observed, an XBox had a 733MHz PIII in it. If we had a slightly more aggressive GPU, I'd be so bold as to remark on us being in the same class as it- but that it fits in the palm of your hand. Since we don't have a handle on what all it'll do, I'm hesitant to say more than a GC in power and ability in the palm of the hand. We'll see once it gets out proper and what all people attempt to put ON it as to where it lies on the power scale.

 

[nikkopt]

(A Core Duo runs faster at lower clock speeds than a P4, for example- or more compelling, an ARM11 machine will run 1/2-1/3 as fast per clock when compared to a Cortex-A8 machine and consume the same power doing it...)

Unless you're talking about a program that is very SIMD heavy then I have a very hard time believing this. I'd like to know how you came up with that number.


yep, those numbers only apply to SIMD performance

edit:

QUOTE
Wow, I expected better performance than a Pentium 3!


I'm guessing that you will install Blender on your Pandrora and use it to render your scenes, right?

 

[WizardStan]

(A Core Duo runs faster at lower clock speeds than a P4, for example- or more compelling, an ARM11 machine will run 1/2-1/3 as fast per clock when compared to a Cortex-A8 machine and consume the same power doing it...)

Unless you're talking about a program that is very SIMD heavy then I have a very hard time believing this. I'd like to know how you came up with that number.

I'm basing this entirely on memory from a tech presentation I attended about 2 year ago, so I may have my facts wrong, but...
x86 being a CISC processor takes multiple clock ticks to execute an instruction. Over the years, they've actually improved some of the algorithms used such that these instructions actually take fewer clock ticks to execute. ie, while a P2 might take on average 3 clock ticks per instruction, a P4 can execute an instruction every 2 clock ticks. Numbers generated by means of anal extraction, but I think you get the point.
I'm not sure whether RISC processors can benefit in the same way.

 

[yashton]

QUOTE
In terms of power most desktop x86s draw around 30 to 60 watts, an ARM is more in the 1 to 4 watt range.


Love to hear when someone actually knows their physics.

Research is your friend

Good article on RISC reasoning and design
http://en.wikipedia.org/wiki/Risc
Basically it revolves around making it easier for a compiler to create machine code than for humans to write assembly.

Cortex-A8 is the processor in the Pandora. It is in a family of processors known as ARM.
http://en.wikipedia.org/wiki/ARM_architecture
http://focus.ti.com/general/docs/genconten...contentId=49075
ARM is used in pretty much everything that isn't a server or a desktop. Most PDA's, Smartphones, and other such devices use it.

Pretty much everything that isn't a desktop PC runs something other than x86. Every current generation gaming console, the Wii, Playstation 3, and even the XBOX 360, run a POWER architecture. For embedded systems and portable devices it's not even a question. x86 is a venerable dinosaur, but it simply has enough inertia in the desktop market, and enough binary software out there that cannot have the compatability broken, that you won't be able to break free from it until the market goes completely open source. Even Apple switched to x86 (and man, I miss PowerPC, still run my Powerbook Aluminum as my laptop).

The one thing that always gets me about this discussion is how people keep bringing up Windows, especially non-games. Why the hell do you want to run Windows applications? Obviously they have a very large significant game library, but for heaven's sake, free yourself from the claws of Microsoft! There are a million games we can emulate and a large collection of open source games to play, and most things you could run with an equivalent Windows desktop there are open source equivalents. Are you expecting to play Half Life 2 or something?

 

[benjiro]

The one thing that always gets me about this discussion is how people keep bringing up Windows, especially non-games. Why the hell do you want to run Windows applications? Obviously they have a very large significant game library, but for heaven's sake, free yourself from the claws of Microsoft! There are a million games we can emulate and a large collection of open source games to play, and most things you could run with an equivalent Windows desktop there are open source equivalents. Are you expecting to play Half Life 2 or something?

This is coming from a guy who's primary os is linux. Unfortunately, not all programs have a "as good" equivalent, as some of the windows programs.

For instance, i have yet to find anything that can replace "comicrack". There are alternatives, but none with the same functionality ( the good news, the developer is looking into mono ). Gimp... *arg*, give me Paint Shop Pro 7 any day. I hate! that the sidebars are not dockable. Yea, yea, they are working on it.

Its definably not easy to switch from Windows to Linux. How many people know that Audacious is a replacement for winamp? Krusader -> Windows Commander. Etc. Problem is, you can tell that Linux software is doing a cache up job, to the user friendliness of a lot of the windows programs.

A client of mine was saying the same thing. They have a IT'er who promotes open software, and they use openOffice etc... But it creates problems with the "outside" world... They forget that they are using the native openoffice file format. Wham. The person they mail, can't open the document files. Etc

This is just my opinion ... I'm not saying, if you exclude gaming, that most people can't switch, but, it takes so much time researching, finding the new programs, finding out that some have no equivalent, etc... For a IT person, sure. But 95% of the planet will not switch because they don't see a added value to switching.

Btw yashton, you forgot to point out how many unix or linux based devices people have in there homes, and don't realize it

 

[WizardStan]

Gimp... *arg*, give me Paint Shop Pro 7 any day.

Could that not just be because you started with PSP and GIMP is just different? If you'd started with GIMP, do you suppose you might have the exact opposite arguments? I've never used PSP, and I think GIMP works just dandy.
You're right, there aren't exact perfect copies of everything for Windows, nor would I want there to be. Different developers take a different path, with different goals, different strengths and weaknesses. Learn to use all the tools across all the media, not just the ones for your choice platform, and you'll never be at a loss.

 

[whiskthecat]

Well if you want to look at MIPS (which aren't really comparable either, but better than comparing mhz), then a P3 has 2.708 MIPS/MHz while the Cortex has 2 MIPS/MHz

So a 500MHz Cortex = 369.276MHz P3
And 900MHz Cortex = 664.697MHz P3 which is about 91% of the aforementioned Xbox

If you wanted to continue the Xbox comparisons....

Pandora Memory 256MB 333MHz vs Xbox 64MB 200MHz

SGX 14 MPolys/sec vs NV2A 125MPolys/sec

But, like I said, I don't claim any of these to be actually worth comparing.

 

[Jonttu]

Well if you want to look at MIPS (which aren't really comparable either, but better than comparing mhz), then a P3 has 2.708 MIPS/MHz while the Cortex has 2 MIPS/MHz

So a 500MHz Cortex = 369.276MHz P3
And 900MHz Cortex = 664.697MHz P3 which is about 91% of the aforementioned Xbox

If you wanted to continue the Xbox comparisons....

Pandora Memory 256MB 333MHz vs Xbox 64MB 200MHz

SGX 14 MPolys/sec vs NV2A 125MPolys/sec

But, like I said, I don't claim any of these to be actually worth comparing.

Basically, guys, stop comparing apples and oranges. MIPS are considerable when comparing proc's, but at the end of the day your just counting how many ships vs. yachts pass the harbor each day.

 

[mali]

I ask myself, how many of you that ask questions about the horsepower of the Pandora and compare it to x86 desktop, have ever used a handheld of a similar category. The day you start to use it, you will change your mind and think differently. It is mindblowing to see a small device like this doing computational tasks without vibrating, heating, noising and gasping for more juice every 2 hours.
You ask yourself, how is that possible, that can't be, where is the cord to the mains?
ARM tries to keep the same power footprint while raising performance, not the other way around like others do.
You won't be able to do some things you do on a desktop, but the things you can do, you do with a grin on your face

 

[icurafu]

I can't think of another CPU that has more bang per watt. Unless a prototype cortexA9 exist somewere.

Bang per watt is exactly what you want in a handheld device.

And the nvidia score was obtained using TnL to repeaqt the same draw triangle. PowerVR could do they same thing if they didn't have an ounce of integrity.

 

[andys]

I'm basing this entirely on memory from a tech presentation I attended about 2 year ago, so I may have my facts wrong, but...
x86 being a CISC processor takes multiple clock ticks to execute an instruction. Over the years, they've actually improved some of the algorithms used such that these instructions actually take fewer clock ticks to execute. ie, while a P2 might take on average 3 clock ticks per instruction, a P4 can execute an instruction every 2 clock ticks. Numbers generated by means of anal extraction, but I think you get the point.
I'm not sure whether RISC processors can benefit in the same way.

Unfortunately it's not *quite* this simple (is anything)?

The X86 has a mix of instructions, some of which will execute in a single cycle (simple ops between two registers, for example), and some of which are more complex (for example, incrementing the value in a memory location). However, modern X86 (Pentium Pro and beyond) use a number of clever techniques to reduce the effect of more expensive instructions. The Cortex is a much simpler architecture in comparison - more akin to the original Pentium.

The Atom is much more similar to the Cortex, but it does have more functional units (faster floating point for example).

A good bit to read about is Out of Order Execution (http://en.wikipedia.org/wiki/Out_of_order_execution) which is the technique used by modern desktop CPUs. The next generation ARM will apparently use the same technique (the Cortex A-9), but that is probably a good 2-3 years off.

The ARM has a number of instructions that take more than a single clock cycle to complete, such as multiply and divide (at least on integers), and it also has a number of fairly unpleasant aspects from a cursory reading, such as a mandatory pipeline stall to do floating point comparisons of at least 20 cycles.

(This is somewhat supposition, but I believe you would have to do an FCMPS, followed by copying the status register back, which is a 20 cycle penalty. Even the NEON equivalent has the same hit, but you can do lots of comparisons in parallel)