What Is The Clock Frequency Of The Pandora ?

[christo930]

I don't get into comparing the mhz ratings across different cpu's, but when it comes to performance, how would a arm 9 at 240mhz compare to a p2 at 240mhz? And how would the SOC compare to a pc running at 240mhz?

Or, more to the point, who cares?


Someone porting a game from the pc to the gp2x or gp32. From my point of view, it's kinda just for trivia. To think I have such and such a power and memory and storage sitting in my pocket running on 2 AA batteries, that was state of the art maybe 10 years ago.

Chris

 

[MWeston]

mweston never said the Pandora will run at 1GHz. He just said something about the Cortex-A8 at 1GHz using a certain ammount of Wattage and comparing that to a RISC based CPU developed by the Chinese for home computers.

He would have got the power consumption figures from the ARM page, even if they were a little low for 1GHz.

...unless CPU = 550MHz; GPU = 120MHz; Memory = 333MHz.

550 + 120 + 333 = about 1GHz.


Oh no, that isn't the case.

Truthfully, the numbers I list are just what I have found to be publicly available. I have confidential information from TI that suggests speeds this processor can run at and it is greater than 550MHz but we have a limited number of development kits and a handful of raw parts for custom hardware so it isn't enough for me to officially say that this chip runs at xxx MHz on all hardware. The fact is, someone at TI has run this chip at 1GHz under circumstances unknown to me.

I imagine when the kernel is very stable and running all peripherals, Squidge is going to blast it up to 1GHz and see what happens. If it works, we know at least one chip out there can do it

Quoting figures of 300mW @ 1GHz was taken directly from the ARM website. I hope people understand that it doesn't mean the OMAP3 chip uses 300mW at that speed. That power does not take into account all the peripherals and probably not the cache itself (since size is variable and ARM site doesn't list power for each possible size their customers will use). It also doesn't account for the advanced power management control instituted by TI.


Finally, my comments on the GP2X @ 250MHz: Actually, the ARM site lists that speed as a maximum for the 0.13um process, but the MMSP2 uses a 0.18um process right? That makes the max speed only 200MHz so the fact that all systems get more than that is a credit to the fab making the chips!

 

[christo930]

I don't get into comparing the mhz ratings across different cpu's, but when it comes to performance, how would a arm 9 at 240mhz compare to a p2 at 240mhz? And how would the SOC compare to a pc running at 240mhz?

Chris

I thought my post made it clear that I aware that comparing mhz to different chips doesn't work. It only works when comparing the mhz ratings of the same exact chip. Like comparing 2 p4's (using the same core) for relative performance. But you can't say a p4 at 1.6ghz is faster than a G5 at 1.5 ghz just because of the mhz rating (when actually G5 smokes the P4 at the same clock speed).

I was just curious as to the relative performance of an arm 9 to a Pentium at similar clock speeds and the relative performance of the SOC vs a PC (from the late P1 era).

 

[Exophase]

I hope people aren't comparing clock speeds across desktop processors with the arm one..

I don't get into comparing the mhz ratings across different cpu's, but when it comes to performance, how would a arm 9 at 240mhz compare to a p2 at 240mhz? And how would the SOC compare to a pc running at 240mhz?

Chris


There are a ton of variables that make it hard to compare, for instance on a Pentium 1 you'll end up with a 60 or 66MHz bus speed (compared to what would be 120MHz for the GP2X), slow EDO SIMMs (can be supplemented with L2 cache on the motherboard), and L1 cache amount/associativity that's the same (but with a better way replacement algorithm). On the other hand, the processor itself is a more sophisticated superscalar design with branch prediction and has an FPU (that's pipelined and has okay performance). The instruction set is arguably worse in general, but has MMX.

Pentium Pro at 200MHz (the fastest they made them) has a CPU that is even better, 256KB of on die (fullspeed) L2 cache but still a 66MHz bus, and it has half the L1 cache (and half the associativity for data, which isn't very good).

Pentium 2 at 233MHz has 512KB of L2 cache running at half the speed, back to 16KB of L1 cache, still only a 66MHz bus.

K6 CPUs around the time would be 66MHz buses and have a lot of L1 cache, although only 2-way set associative.

I think pretty much any CPU with a clock around that would be more advanced than the one in the GP2X, but the memory might play out a little better in the GP2X's favor. However, some of them were starting to have more L1 cache or L2 cache moved internally which would benefit a lot more than the better memory.

In general you can probably expect the GP2X to perform noticeably worse, especially if you're porting something with floating point code.

 

[whiskthecat]

It also doesn't account for the advanced power management control instituted by TI.

That's exactly what I was getting at in the other thread.

 

[fusion_power]

It's a surprise.

Especially for the Batteries. :lol:

 

[icurafu]

P1 didn't have MMX. They released a Pentium MMX later for the mainstream while Pentium Pro was being snapped up by performance demanding individuals.

I was the poor idiot who had to settle for the Cyrix 6x86 200.


In terms of DMIPS, the ARM9@240 is 264 DMIPS.

The Pandora at 1GHz is 2000 DMIPS.

From what I'm told PentiumIII@833 has 2933 DMIPS and Pentium M 940 (1.73) gets 7400. Divide each of these by power consumption and the CortexA8 is the outright winner.

What's interesting is that the new Silverthorne at 600MHz and 900MHz will score lower than the CortexA8 at 1GHz. Making it perform slower and use more energy. But intel are on the way and have come a long way.

 

[Pickle]

It's a surprise.

1 Mhz!

reminds me of one of those games where you guess how many gum balls are in a clear bucket

 

[Sphinxter]

What's the frequency Kenneth?

 

[Exophase]

P1 didn't have MMX. They released a Pentium MMX later for the mainstream while Pentium Pro was being snapped up by performance demanding individuals.

Any Pentium 1 CPU around 240MHz would be a Pentium with MMX. That is the CPU I was referring to.

I was the poor idiot who had to settle for the Cyrix 6x86 200.

Probably not that bad if you didn't care about floating point performance.

In terms of DMIPS, the ARM9@240 is 264 DMIPS.

The Pandora at 1GHz is 2000 DMIPS.

From what I'm told PentiumIII@833 has 2933 DMIPS and Pentium M 940 (1.73) gets 7400. Divide each of these by power consumption and the CortexA8 is the outright winner.

Every single one of those numbers looks way too even (and sounds like something published). Some actual results of benchmarks not done by the ones selling the CPU would be much more interesting to me. I know for sure the first two are from ARM's claims of DMIPS/MHz are 1.1 and 2.0 respectively for ARM9 and Cortex-A8, and I also know that the only Dhrystone benches I've seen for GP2X have given it more like 170 DMIPS (granted, GCC sucks at ARM).

What's interesting is that the new Silverthorne at 600MHz and 900MHz will score lower than the CortexA8 at 1GHz. Making it perform slower and use more energy. But intel are on the way and have come a long way.

Where are you getting this information? I'm very skeptical that Intel will be pushing an x86 CPU that performs at half the IPC of Pentium M (according to you), even if it's specifically low power. They're not VIA.

 

[senquack]

I was the poor idiot who had to settle for the Cyrix 6x86 200.

Those were a great deal at the time!

 

[icurafu]

Every single one of those numbers looks way too even (and sounds like something published). Some actual results of benchmarks not done by the ones selling the CPU would be much more interesting to me. I know for sure the first two are from ARM's claims of DMIPS/MHz are 1.1 and 2.0 respectively for ARM9 and Cortex-A8, and I also know that the only Dhrystone benches I've seen for GP2X have given it more like 170 DMIPS (granted, GCC sucks at ARM).

Yes, I got the DMIPS from multiplying 1.1. I don't remember anyone ever running DMIPS on the GP2X.

What's interesting is that the new Silverthorne at 600MHz and 900MHz will score lower than the CortexA8 at 1GHz. Making it perform slower and use more energy. But intel are on the way and have come a long way.

Where are you getting this information? I'm very skeptical that Intel will be pushing an x86 CPU that performs at half the IPC of Pentium M (according to you), even if it's specifically low power. They're not VIA.


Yes, I got the DMIPS from multiplying 1.1. I don't remember anyone ever running DMIPS on the GP2X.

A number of places discuss it now. It started from this article I think...

http://arstechnica.com/news.ars/post/20071...processors.html

 

[christo930]

P1 didn't have MMX. They released a Pentium MMX later for the mainstream while Pentium Pro was being snapped up by performance demanding individuals.

I was the poor idiot who had to settle for the Cyrix 6x86 200.

In terms of DMIPS, the ARM9@240 is 264 DMIPS.

The Pandora at 1GHz is 2000 DMIPS.

From what I'm told PentiumIII@833 has 2933 DMIPS and Pentium M 940 (1.73) gets 7400. Divide each of these by power consumption and the CortexA8 is the outright winner.

What's interesting is that the new Silverthorne at 600MHz and 900MHz will score lower than the CortexA8 at 1GHz. Making it perform slower and use more energy. But intel are on the way and have come a long way.



From what I can recall (and granted I am getting older) the Pentium Pro was a disaster for Intel because Intel assumed that the world would be using pure 32 bit software by the time it was released, but the reality was that windows 9x had a TON of 16bit legacy code in it that would make the Pentium Pro run slower than a P1 at the same clock speed when running Windows 9x. Win NT on the other hand was faster when running on a P-Pro. So users running Windows NT (which were very few at the time) would benefit from the p-pro while everyone else would go slower. So Intel was left holding a new processor, that for most people, was slower then the old one. So in what ended up as a marketing genius, Intel increased the internal cache of the P1 which gave it a 15% increase in speed over the old core and added MMX which would supposedly speed up multi-media (which was a big buzz word at the time), but only when written specifically for it. They used that to buy some engineering time and re-engineered the P-Pro and came up with the Pentium II (which was based on the P-pro) which is much faster than a P1 running legacy code, is on a 100mhz bus, larger more intelligent cache... Those were bad times at Intel.

Only with the release of the CORE architecture has an X86 chip from Intel taken such a big step ahead in terms of performance. In my experience, they have not made such a big jump since the jump to the 486 from the 386. From what I can recall the 486 was more than twice as fast as a 386 at the same clock speed. Lot of people were complaining at the time though, that the 486 was nothing more than a dressed up 386 with a co-processor thrown in for good measure, but the reality was that performance more than doubled.

Chris

 

[whiskthecat]

http://en.wikipedia.org/wiki/Million_instructions_per_second
Check out that chart, the cortex is listed, but Wikipedia says that you can't really compare MIPS between architectures.

 

[icurafu]

I read that DMIPS is more comparable. Which is what I used.

 

[Exophase]

From what I can recall (and granted I am getting older) the Pentium Pro was a disaster for Intel because Intel assumed that the world would be using pure 32 bit software by the time it was released, but the reality was that windows 9x had a TON of 16bit legacy code in it that would make the Pentium Pro run slower than a P1 at the same clock speed when running Windows 9x. Win NT on the other hand was faster when running on a P-Pro. So users running Windows NT (which were very few at the time) would benefit from the p-pro while everyone else would go slower. So Intel was left holding a new processor, that for most people, was slower then the old one.

I doubt the problem was Windows nearly as much as it was the legacy software that was still being run on it. Of course I also doubt Intel ever really meant for Pentium Pro to be a mainstream commercial chip either, given that they didn't add a lot of onboard L2 cache again in a mainstream CPU until the middle of Pentium 3's life cycle. I see it as more of a precursor to the Xeon line. Until this point Intel really didn't have a separate line of server grade x86 CPUs.

So in what ended up as a marketing genius, Intel increased the internal cache of the P1 which gave it a 15% increase in speed over the old core and added MMX which would supposedly speed up multi-media (which was a big buzz word at the time), but only when written specifically for it. They used that to buy some engineering time and re-engineered the P-Pro and came up with the Pentium II (which was based on the P-pro) which is much faster than a P1 running legacy code, is on a 100mhz bus, larger more intelligent cache... Those were bad times at Intel.

Pentium 2 wasn't on a 100MHz FSB until the middle of its life cycle either, and the only thing Intel changed to improve legacy performance was adding segment register caching, which I suppose made a big enough difference (286+ protected mode uses segment registers in a totally different way than real/virtual mode so the architectural implementation is quite different, I take it that the original P6 was converting between the two on the fly every time).

Only with the release of the CORE architecture has an X86 chip from Intel taken such a big step ahead in terms of performance. In my experience, they have not made such a big jump since the jump to the 486 from the 386. From what I can recall the 486 was more than twice as fast as a 386 at the same clock speed. Lot of people were complaining at the time though, that the 486 was nothing more than a dressed up 386 with a co-processor thrown in for good measure, but the reality was that performance more than doubled.

I don't think this is a fair comparison because, like many people, you seem to be ignoring clock cycle time improvements altogether which is actually much worse than ignoring instructions per cycle in the long run. If you took a 386 like architecture I doubt you'd be able to scale it past a few hundred MHz, even using the best technology available today. Pentium 4's problem was that it released at too low of a clock speed and Intel made some assumptions that led to what be called design mistakes (some of which were corrected), but in the end it improved performance probably about as much as any other generation gap if you compare the fastest out at release to the fastest out at release of the next generation. Only now have they hit other problems with increasing clock speed so much that's inhibiting this approach and requiring the design of more "wide" architectures (and they'll exhaust instruction level parallelism soon, so it'll be on to nothing but lots and lots of cores and inflated performance numbers from these)

QUOTE

http://en.wikipedia.org/wiki/Million_instructions_per_second
Check out that chart, the cortex is listed, but Wikipedia says that you can't really compare MIPS between architectures.


MIPS aren't DMIPS. DMIPS (Dhrystone MIPS) are the result of a Dhrystone benchmark normalized against a reference machine that's supposed to be 1MIPS/MHz (but is actually more like half that, so the numbers are way off, but the instruction set is radically different anyway). Unfortunately, nowhere has ARM said that Cortex-A8 can achieve 2000 MIPS, just DMIPS, which tend to be higher than MIPS thanks to the normalization number being so off in the first place. ARM actually claimed that after testing dozens of industry standard programs it only achieves 0.9 IPC, which would be 900MIPS at 1GHz, but I can't really reconcile with how this number can be so low given the architecture, unless none of these programs were optimized for Cortex (even then seems way too low). I'd like to see what ARM claims for previous generations though.

The reality is that if you have an n-issue CPU you can almost always engineer a benchmark to run at very close to n instructions per cycle just by taking a very long loop and filling it with instructions that schedule well to make full use of the pipes/execution units. Unfortunately highly superscalar CPUs can often issue more instructions at a time than typical code will be able to fill in the backend (not enough execution units) and memory, branches, and dependency chains, just to name a few things, will cause lots of stalls. So I don't trust MIPS ratings, not so much because of the ISA being different between CPUs, but because the benchmarks could be totally unrealistic. I'm really skeptical of a benchmark that gives close to three IPC for P6.

I'm probably boring everyone with things they already know though ;P Oh well, what goes around I should edit that wikipedia node sometime >_>

 

[Exophase]

Yes, I got the DMIPS from multiplying 1.1. I don't remember anyone ever running DMIPS on the GP2X.

Dhrystone benches by dwelch (they're about 1.5 years old): http://www.dwelch.com/gp2x/

Interestingly he says that ARM advertised 0.95 MIPS/MHz for ARM9, not 1.1. The actual result he got was closer to 180 than 170, my mistake there.

A number of places discuss it now. It started from this article I think...

http://arstechnica.com/news.ars/post/20071...processors.html



That's really interesting, that Intel might end up sacrificing architectural advancements for power consumption just to come up below ARM on both counts. What's more, ARM is rolling out Cortex-A9 which will be out of order. Given the specs it'd make sense for Intel to pattern the pipeline off of Pentium 1 so that current compilers could optimize tightly for it. I wouldn't write it off entirely though, it's probably meant to compete against VIA's C7-M which is only single issue.

On the one hand, it makes it seem pretty odd that Intel abandoned XScale. Do people want x86 that badly in their handhelds?

 

[icurafu]

All I can think of is Lindows. X86 would still be useful for running existing x86 apps, considering intels long term plan. (mooretown - the 24 hour battery chipset) which will no longer support Windows OSs.

So it does seem that this is there answer to ARM.

 

[javaJake]

It's a surprise.

 

[Karel Jansens]

From what I can recall (and granted I am getting older) the Pentium Pro was a disaster for Intel because Intel assumed that the world would be using pure 32 bit software by the time it was released, but the reality was that windows 9x had a TON of 16bit legacy code in it that would make the Pentium Pro run slower than a P1 at the same clock speed when running Windows 9x. Win NT on the other hand was faster when running on a P-Pro. So users running Windows NT (which were very few at the time) would benefit from the p-pro while everyone else would go slower.

OS/2 kicked [NT's] butt on a PPro!