Pandora outdated?

[timbur666]

Can't you just build a giant castle with the leftover OMAP chips and fast-forward production to DM3730?

Yes, with a donation of 50,000 USD, I could.

ED, if only you were writing this in German notation, where the "," is the radix point! I could afford a $50.000 donation, just not a $50,000 donation.

 

[theshining]

Nice thread ED, sounds like it was started to stop a bit of potential trolling, now where did I put my 18 megapixel compact camera? I have to take a piccie of a pair of used socks for e-bay!!!

 

[dementio]

Nice read, actually learned some stuff


But I still wanna try upgrading the SoC on a spare OP board if I can ever get my hands on one (just `cause)

 

[Grench]

BTW: While the Pandora SoC has a standard clockrate of 600MHz, actually, quite a few can clock up to 1,2GHz, a lot to 900 - 1GHz and most to 850MHz. If you check the web, the current DM3730 is being sold as 1GHz version.


What's the difference to the old DM3730 or our Pandora CPU? Basically nothing.

I thought the DM3730 was on a smaller lithographic process - with a stock clock of 1Ghz. Comparing it to the 500Mhz CPU overclocked to hell and back isn't overly fair. How far will the DM3730 overclock to with stability?


I know - it's been said in these forums before. We need to use up the existing CPU supply. It's just hard to justify buying a 2nd Pandora unless it's improved over the one I already have.

 

[milkshake]

just wait untill the current soc surply runs out, ED already said when that happens he will be buying DM3730 as they are cheaper than the current SOC

 

[Exophase]

Well, I knew the Tegra 2 didn't use it, and according to the ARM information page about the Cortex-A9 stating "Optional NEON™ media and/or floating point processing engine", I thought there were more SoCs that do not have NEON included, especially some in these cheap chinese devices that will probably keep flooding in.

Correct, it could have been missing on any Cortex-A9 SoC, but it's only Tegra 2 (cheap ones from Chinese manufacturers like Amlogic and Rockchip have it). You barely save die space by not having it on a dual or especially single core SoC, it's really tiny.. at 40nm like Tegra 2 they probably saved at most 1-2mm^2 (whole chip is 49mm^2). I suspect that nVidia made this decision for marketing/strategic reasons, or were just completely mistaken somehow. Tegra 3 corrects this and adds NEON.


This makes a compelling argument for comparing against a Tegra 2, perhaps. But Tegra 2 is barely used in phones and is being phased out in tablets. OPT would be crazy to use a Tegra 2 in a device that's meant to be compatible with Pandora, so I'd rather take it off the table completely for this discussion.

The benchmark wasn't done on a smartphone. I got this directly from an SoC manufacturer, who made a direct speed comparison of the A8 and A9 (at same CPU speed).
The blog post I linked to stated the same, so I'm pretty sure this is accurate.


This is just a stock CPU comparison though, not including any optional stuff.

Vendors and various third parties are pretty bad at benchmarking too. In the former case companies can be especially misleading because they always have something to sell you (even TI knows you can buy OMAP3530 or DM3730 but not OMAP4s). It just varies way too much to make a broad statement, but I think any 10-20% figure is absorbing a lot of cases were the speed isn't improved much because things are dominated by memory performance or a somewhat atypical case where the scheduling on A8 is as good as it'll be on A9 and there aren't a lot of dynamic stalls. Highly optimized NEON code can actually run faster on A8 clock for clock. But I still wouldn't trade that vs an A9.


In the case listed you do have an A9 vs an A8, but it's a pretty high-end well done A8 implementation on 45nm vs a crappy 65nm budget A9. Samsung SoCs for instance have much better memory bandwidth than a lot of the competition, and in this case it has much more cache. So yes, you have to consider more than just the CPU core. But with everything else equal A9 will often win a lot. Maybe that's the point you were really trying to get at.

How much difference can there be, speedwise?

Varies way too much per program to really make a statement, and I'm not aware of any tests on these platforms.. main memory performance makes a huge difference too (more cache helps more if your memory is slower). For some data points I'd look at current Celeron and Pentium vs higher end Intel chips that have more L3 cache.

Yeah, I probably didn't make myself clear with the post.


Specwise, this COULD make a difference - but I was looking at it from a realistic view.


How realistic is it that many application actually make use of that?


N64 emulation would be faster anyway (I mentioned that above as well), but how realistic is it that somebody would port stuff like DS or PSP emulation working fullspeed on a Cortex-A9 right now?


Apps ARE slowly changing to make use both cores, but I'm not sure how fast this change will really happen (took ages on a PC).

It's way more realistic that programs will benefit from multi-threading than that'll they use the DSP, or even make good use of NEON. In the end for something like Pandora you benefit a ton based on what a handful of programs do, not what everyone is doing on average. A few killer app that benefits from multiple cores is enough to justify it. I don't know about it helping DS or PSP emulation ports, but the best emulators on Pandora aren't just ports of PC emulators. So why should we use that standard when thinking of possibilities for new emulators?

I've got a BeagleBoard-xM, and Ari also made a post about how he tried running N64 on the DM3730, didn't give a huge speedboost.


According to various posts I found on the internet, the DM3730 doesn't go higher than 1,2GHz when overclocking.


I'd love to do more tests on that myself, but that's what I found so far.


The BeagleBoard-xM doesn't feel faster for me than my Pandora does (running ArchLinux ARM on both as a test).

With all of the problems with say, Pandaboard's performance, I put little stock in these things being configured right. And a subjective test about how fast something "feels" isn't really that useful. It doesn't make any sense for DM3730 to not be faster, and TI wouldn't have added things to it if it did nothing.

As mentioned above: I'm seeing it in a realistic way, not pure technical (should've made myself more clear about that).


People even stated in this post that they didn't even use the GPU at all until now.


Why would they do that on an A9 suddenly?

If Android were used that GPU would suddenly become much more relevant. That's the gateway, more than the CPU type. But even with the kind of Linux Pandora has if the GPU was better there might be more porting efforts. And assuming Pandora had better manufacturing capacity it'd sell more with better specs (and/or cheaper price..) and that'd attract more development. It's not really fair to say that no one cares about GPU when current circumstances and specs are holding back interest.

I'm not sure others really do that. Every Android phone I tried (tried Samsung Galaxy S2 and the likes as well) felt somewhat sluggish.


I was looking for a new phone, so I tried various phones, but in the end, I bought a used N900 (had one already).


The only smartphone OS that doesn't seem laggish to me (at least in the UI) is the iPhone.


But that's probably because the OS is optimized to run on one single hardware with huge money behind it...

UI response time is a software issue, but that doesn't mean that critical loops in games and emulators are going to all around run slower on these devices. The two things are pretty separate. You can't really compare how different UIs feel to judge hardware or other software quality.

And still, Forget-Me-Not on an iPhone 4 needs more frameskip than it does on the Pandora.
The iPhone4 is also still using an Cortex-A8. Would anybody say the iPhone4 is outdated? (Well, it sure is, as the 4s is newer, but who would SAY that the iPhone4 is outdated?)


Are there any games that really NEED an iPhone4s and wouldn't work on an iPhone4?


HDMI out is the only real thing missing here, not sure we need 1080p on an LCD


But yeah, if you're using it as a media player, it would be nice.

Yes, I'm pretty sure people would actually say 4 is outdated compared to 4S. Of course, iPhone 4's hardware is better than Pandora's too.. I don't think games need 4, but it's kind of hard to bridge this with Pandora when a vast majority of phone games aren't available for it. With most games being PC ports it benefits from CPU more there. I don't know about Forget-Me-Not but most games available on the app store will have enough development effort to not end up behind a Pandora port in performance.


Yes, 1080p is actually nice to have when you also have good TV out.

True, with 800MHz, it won't be 8 times as fast.
Though I remember reading somewhere that the A15 will have about 4x the performance of the A9 with the same speed.


And as the A15 will start at 1,5 - 1,8GHz, it will have about 8x the performance compared to our A8 with 800MHz.

Take what you read with a little more scrutiny. You actually thought that an A15 could improve per-clock per-core performance over A9 4 times (and you think A9 is effectively zero faster than A8)? It'd be substantially faster per-clock than a Sandy Bridge. The real world performance improvement per-clock is going to be more like 1.5x. I would give Cortex-A9 a typical 1.2x improvement over Cortex-A8. Cortex-A15 in a Pandora like device might clock at 1.8GHz but it'd positively chew battery, still: (1.8 / 0.8) * 1.2 * 1.5 = 4.05x. A 1.2GHz Cortex-A9 that can be had today would give you about 1.8x, which is still substantial.

True. But what WILL be realistically be done?
I'm sure the Pandora would be capable of running PS2-Like games, if there would be somebody who coded it.


I think we can agree we cannot expect big blockbuster commercial titles on the Pandora, so we'll most probably never see a game like that, but it would be possible.


So regardless of having an A8 or A9, we probably wouldn't see any other games.


About compilation: Yeah, that would be neat. Though I'm not sure how many users are actually compiling on their Pandora.

The way I see it, the more you limit Pandora to its CURRENT niche the more you DO want CPU because you're doing more desktop-like things instead of phone-like things. But are you saying you're against Android, or..?


As for compilation, I don't know, how many Pandora owners are developers.. a lot? Maybe I'd compile on it if it were a lot faster

Sure, whatever the OS allows is within warranty.
OPP5 is within TIs specs, and that's the maximum the stock OS allows.


notaz is currently working on Kernel 3.2, which adds some OMAP stuff following TI's specifications.


The kernel forces OPP4 on >=550MHz, OPP5 on 600MHz.

So I guess the stance on that changed at some point. The changes sound good.

Well, I wanted to take a more realistic approach.

Is there ANYTHING you can do on a modern smartphone, that you can't do on the Pandora?


(not theoretically, but realistically).


Emulators like PS2 or GameCube are not possible at playable with an A9 (not sure if that would even be possible with an A15)


I haven't seen a game on iOS / Android that I think wouldn't be possible on the Pandora as well.


The OS itself on phones is needing a lot more CPU power (which makes sense, since it's a phone and should react to phone calls quickly), so some games (like Forget-Me-Not) even run slower on higher specced phones than on the Pandora.


So while the specs are outdated, the system itself isn't in my opinion.


That's what my approach of this was supposed to be.


The A9 is not the holy grail speed wise. The speed boost from the A9 to the A15 will be A LOT higher than from the A8 to the A9.

Clock boosts are the biggest driver in performance right now. A9s are already out that can clock to 1.5GHz on tablets. I don't know what a Pandora can do; it's a lot bulkier and has a much bigger battery than a phone. But it's not a tablet.


hlide has said he believes in PSP emulation on a 1+GHz dual core A9 with NEON. I believe in DS emulation on similar hardware, or at least much better DS emulation than possible on Pandora. Not with ports of anything though. And I still don't really buy into the "OS is slower and makes your CPU effectively slower" mindset, I think the comparison is way more complex than that.

Nope, as said, I am no speed whore.
I want to do a proper comparison in a realistical way, not theoretical way.


No app or game currently running on an A9 will probably NEED one. It would work fine on an A8 as well (might need some optimization, but should run).


So when is a hardware outdated? When new hardware is out, or when it's so slow it can't compete with current systems (not based on specs but on user experience)?

My biggest problem with your viewpoint is your black and white determination of if hardware is "needed" or not. If you want realistic, why not ask the users, are there times when they wish the Pandora was faster? Then it'd benefit from being faster. Maybe ask developers too. Limiting everything to PS2 and Gamecube emulation is way too narrow (and do you think an A15 SoC will suddenly have those emulators at "good enough"?) Forget DS and PSP, what about Saturn? Or less known platforms like PC-FX? What about web browsing.. flash, HTML5? What about boot up time? What about battery life? I don't really like these "everything is good enough and therefore can't be better" mindsets. But it's not like it's the first time we've had this argument so I'm just going to leave it at that.


And I don't like the bit about other phones being saddled by their OSes so therefore Pandora isn't outdated. It's not a phone, it doesn't have the app catalog, it's apples to apples. Pandora's hardware IS outdated because better hardware can be bought for the same price. That's a really critical evaluation point. It has been passed up by multiple design AND process revision stages (which are usually not done at the same time). I'd feel better if the question was instead "is Pandora's hardware good enough for your needs despite being outdated?"


I agree with you in a sense, though, that it's the specs that are outdated rather than the Pandora.. but that's more because the Pandora hasn't been updated nor replaced with any real competitor. I worry that lack of competition results in stagnation (but also opens up a lot of hurt for if a competitor really does drop down)

Don't be confused!
That was Michael saying that, not me. I'm still planning to use the DM3730 in future Pandoras.


Maybe I made myself a bit more clear what my goal was... not easy if English is not your native language.

MWeston said that you had to use up the OMAP3530s, but don't you agree with this? Still, in your opening thread you say the chip is the same. Why bother? Or is it just because OMAP3530 is eventually going to be EOLed? If it isn't already..

 

[canseco]

Talking about N900.


To run Gran Turismo 2 at full speed (and still with little stuttering) i need to use dsp kernel profile (kernel power kp49) and overclock cpu to 850 MHz.


Dsp overclock is experimental, and crashes in minutes, and temperature goes 42 or more.


I used same emulator as Pandora have.

 

[MrChops]

Can't you just build a giant castle with the leftover OMAP chips and fast-forward production to DM3730?

Yes, with a donation of 50,000 USD, I could.

Potential buyers for your stack of omap3550's:


http://www.embedinfo.com/


http://opensourcemid.org/


http://www.seeddsp.com/ProductView.php?Viewid=14&p_classid=61


http://www.witstech.net/wits_a81hplus.html


http://beagleboard.org/


You have about 2000 left, so you need $25 a piece for them. I have no idea if that's a good price but, you say they are more expensive now to buy, perhaps someone is in need of 2000 at the old price?

 

[EvilDragon]

MWeston said that you had to use up the OMAP3530s, but don't you agree with this? Still, in your opening thread you say the chip is the same. Why bother? Or is it just because OMAP3530 is eventually going to be EOLed? If it isn't already..

TI has a 10+ years life for the OMAPs, so the 3530 will be available from TI until 2018.


There are still a lot of applications using the 3530.


Sure we've got to use up the OMAP3530, but it's 2000 we have.


After that, the DM3730 could be used.

 

[Grench]

MWeston said that you had to use up the OMAP3530s, but don't you agree with this? Still, in your opening thread you say the chip is the same. Why bother? Or is it just because OMAP3530 is eventually going to be EOLed? If it isn't already..

TI has a 10+ years life for the OMAPs, so the 3530 will be available from TI until 2018.


There are still a lot of applications using the 3530.


Sure we've got to use up the OMAP3530, but it's 2000 we have.


After that, the DM3730 could be used.

Some considerations that might allow shortening this process...


A. What did the OMAP3530 cost per CPU?


B. If you were to sell off 1000 of the OMAP3530 CPUs in bulk - what could you get for them per unit?


C. How much would it cost per unit for a 1000 piece order of DM3730 CPUs?


D = (A-B+C)*0.25 (adjustable based on your desired markup for this component)


E = 375EUR (current Pandora listed price) + D


E would be how much to charge for a, "Pandora 1.5" with 512MB and a DM3730.


-Your existing customers would have a new model to purchase (see repeat customer).


-Your existing pre-orders would be able to buy up to the newer model - if they so chose. Since they would be at a higher cost, it would help get the old pre-orders out quicker.


-The existing model would still be nice in it's own right. We're not talking about a whole newly designed machine.


-You would simply be offering the same computer with two processor options.


-You could offer an upgrade-for-a-fee for Pandora consoles already in the wild. This would essentially be making the same income as a unit sale without needing to inventory additional plastic parts.


I'm probably simplifying it too much though. I'm sure there is more to it than that. The initial DM3730 purchase cost would be more capital outlay - and there are politics and sentiment to account for. Still, the chance to have existing customers have something to make them repeat customers is, in my humble opinion, a pretty big deal.


Just my humble opinion. How you run this business is, of course, to your own devices.


Doubling RAM is tempting for me to buy a 2nd (new) Pandora (repeat customer). The newer and faster (with overclock) CPU and RAM together - that makes the temptation exponentially higher for me to purchase a 2nd one or an upgraded/replacement motherboard.


Grench

 

[dgame]

I think the key here is to get the people the Pandoras that they have already ordered.


If OPT only has 2000 of the existing CPUs they should be able to use them all up fulfilling existing pre-orders.


Then OPT can switch to whatever they want.

 

[ZXDunny]

Sorry, have to disagree. Multi-threading makes all the difference to performance of many apps if they're written with that in mind. My own PandaBAS performs massively slower on the Pandora at 1GHz than it does on my x86 dual-core 1GHz (old) laptop, and that's purely because it's split into two threads - one for the interpreter, one to update the graphical display. The Pandora has to do both on one CPU and it's slow.

Please, don't compare an x86 with an ARM.


Have you tried running PandaBAS at 1GHz on a single-core x86?


It will probably perform better as well

Yes, I have - a Samsung NC10 netbook, and the performance is noticeably better than on the Pandora. I admit that FPC's code creation is probably very much subpar when compared to the x86 FPC and Delphi, but it can't all be down to the compiler... Still, PandaBAS runs faster at 600MHz than Sinclair BASIC did at 3.5MHz, so I'm reasonably happy


D.

 

[Exophase]

NC10 is running the usual 1.6GHz Atom, are you sure you're running it only 1GHz? That netbook is single core too, although it has hyper threading so you do benefit from threading. That should be an even stronger case for what two real cores can do; you should compare with and without HT enabled if you can.

 

[ZXDunny]

I heard that Forget-me-not runs slower on the iPhone4 than it runs on the Pandora even though the iPhone4 has a faster CPU (and a proper one as well, no cheap one).

I'll have to correct you on this one also


Forget-Me-Not needs frameskip (it's in my next update) on the Pandora. On the iPhone 3 and the iPod Touch, it doesn't need it, and indeed has no code for it. I had to add it because the Pandora cannot keep up. The iPods with the same (or larger) screens as the Pandora run faster also, not needing any frameskip. I've had to make some major changes to the code just to get it running acceptably on our device.


D.

NC10 is running the usual 1.6GHz Atom, are you sure you're running it only 1GHz? That netbook is single core too, although it has hyper threading so you do benefit from threading. That should be an even stronger case for what two real cores can do; you should compare with and without HT enabled if you can.

Ah, I'll give it a go and see if I can disable it, thanks. I didn't know it did hyper-threading.


D.

 

[renegadechic]

"is Pandora's hardware good enough for your needs despite being outdated?"

For me that's pretty much the biggest factor in what ED was trying to say with this post and probably the biggest thing that should matter to anyone looking to buy a Pandora. My big problem with reviewers and tech blogs these days is that they seem far more interested in numbers and form factors than being people who have the facility to delve into what a device can do. They're ultimately big consumers and trying to prove their own relevance by showing off whatever is shiniest or latest or most sought after. Give someone used to smartphones, popular brand handhelds or PCs a completely open linux device, they're pretty likely to scratch their heads and not even know how to approach getting the most out of it. Also, because its less mainstream it isn't going to draw much popular interest and site hits. I am pretty sure its seen as those guys that do that thing they totally agree with and thing is great and all but ultimately John in the next cubicle has an iPhone so why should they fart around with a Pandora.


At this stage, the Pandora as a device is probably only half the story. The big thing is the work being done for the Pandora and the way that progress here spills out into the rest of the world. Lots of apps and emulators have benefited from this work but the delays for the hardware and the numbers game end up overshadowing all of it. Hopefully that will change very soon with regular production. I definitely believe in what's going on with the Pandora and what its trying to do but its a shame we have to try and defend it every single of the way to anyone who doesn't get it. Its something that seems to have dogged the project for its entire lifespan and been resisted, lets not let it creep in and ruffle our feathers because it wont achieve anything.

 

[Link]

@Dunny & ED,


Sorry for any confusion I created with my Forget-Me-Not thoughts. Just playing it on my Dad's iPhone, not sure if its a 4, or 4S, it was slow compared to MY current build on the Pandora which is using frameskip which Dunny gave to me for Beta testing a few months back. Wow, has it been that long? Didn't mean to create any confusion here, but I was just pointing out an observation between how fast the beta is on the Pandora and how slow the iPhone 4? is in comparison, minus frameskip. I believe Nyarlu is thinking about adapting frameskip into the iPhone4 build, but there is only a beta for the iPhone 3.1.? which I'm waiting on my test build currently as there was 0 support for it a week or two ago, unless I'm totally out of the loop.

 

[Stephane Hockenhull]

I could really use a dual-core in my game, the audio decompression and mixer runs on separate threads.

Sounds like a perfect candidate to experiment with the DSP.

yeah, I guess I could play with that.


altho its just around 12% cpu time for a game, moving data in and out of the DSP may take just as much CPU time for synchronisation.


plus the extra effort needed to rewrite the code in assembler for an obscure DSP, and have ppl enable the DSP at the OS level first and reserve memory for it and so on ...


hardly worth the time.


don't forget the code also has to be DEBUGGED! not an easy thing to do with a weird coprocessor.


same problem that happened with Saturn and Jaguar, multiple odd processors = too much work.


uniform multi-processing is much better than specialized DSPs.

 

[Exophase]

yeah, I guess I could play with that.
altho its just around 12% cpu time for a game, moving data in and out of the DSP may take just as much CPU time for synchronisation.

Probably not much more than between dual cores, although you can't go straight through a shared cache. But audio code shouldn't need a lot of synchronization, at least not if just for music, it's pretty fire and forget.

plus the extra effort needed to rewrite the code in assembler for an obscure DSP, and have ppl enable the DSP at the OS level first and reserve memory for it and so on ...
hardly worth the time.


don't forget the code also has to be DEBUGGED! not an easy thing to do with a weird coprocessor.


same problem that happened with Saturn and Jaguar, multiple odd processors = too much work.

You might not know this but this DSP family is available in many TI chips. Fortunately TI provides a compiler for it for free. Trust me when I say you don't want to hand write C6x ASM if you can avoid it. It's very difficult: 8-way superscalar with many variable latencies and no interlocks.

uniform multi-processing is much better than specialized DSPs.

Very much so. In this DSP's case you take a big performance penalty by even writing code that's interruptable.. such that it's often better to write kernels that run several iterations with interrupts off :/

 

[EvilDragon]

I'm probably simplifying it too much though. I'm sure there is more to it than that. The initial DM3730 purchase cost would be more capital outlay - and there are politics and sentiment to account for. Still, the chance to have existing customers have something to make them repeat customers is, in my humble opinion, a pretty big deal.

Yep, sadly, it's not that simple.


First off, nobody will buy the old OMAPs. They've been lying around for about a year now, and although they're good, nobody will take the risk (unless you sell them for a way too low price).


Next, the DM3730 has a leadtime of about 6 weeks.


Meaning: If I order and pay it now, I won't see it before 6 weeks.


However, I'd need to find 2000 customers who would want to pay about 24 EUR more right away, which would probably take another month.


So it would delay everything another 4 months from now on...


I doubt that's a good thing

 

[CAZMIESTER]

Maybe what the pandora needs is a bit more choice as far as os is concerned how about android id pay to have a good build on the pandora not a homebrew build an official you say the pandora is not outdated and is open so why not have a choice as far as the operating system is concerned before every one jumps at this with debian ubuntu etc etc i mean an official build that can be rooted to the marketplace.