The Sony PS3 thread (apparently *still* obnoxious!)
- teamonkey
- Sheepie Whore

- Posts: 12082
- Joined: Thu Jun 13, 2002 1:07 am 0
- XBL Gamertag: itsthetaste
- PSN: teamonkey_ca
- Contact:
There was an academic paper posted on here a while ago that managed to get Cell running near its peak theoretical limit.Kemper_Boyd wrote:I'm sure IBM can give you a power virus if you want to see Cell running
at 100%.
- Kemper_Boyd
- Sheepie Whore

- Posts: 11967
- Joined: Wed Jul 31, 2002 9:15 pm 0
Sure I've read that, did you notice all my comments?teamonkey wrote:There was an academic paper posted on here a while ago that managed to get Cell running near its peak theoretical limit.Kemper_Boyd wrote:I'm sure IBM can give you a power virus if you want to see Cell running
at 100%.
Utilisation isn't a good measure of performance.
A power virus is an interesting example, they really work the chip but in
many cases that work is utterly pointless. I'm sure I've read that
IBM have code for some POWER chips that leads to greater heating
(a good proxy for activity in the frontend, ALUs, register file and caches)
than any code found in shipping operating systems or applications.
Measure performance doing useful work, utilisation is about as helpful
a comparison as MHz or MIPS. i.e. It's not.
- Kemper_Boyd
- Sheepie Whore

- Posts: 11967
- Joined: Wed Jul 31, 2002 9:15 pm 0
I may be misremembering a bit, but I dug this up. It was from an article
about IBM's POWER5, but the figures were from Intel and presented at
IDF in 2004.
So teamonkey even that paper you mention may not be an example of
Cell running close to "100%" i.e. Peak FLOPS isn't necessarily the greatest
workload.
Still I'm sure I've had a conversation about this, about 3-4 years
back arguing the toss about future CPUs for the Mac, and was told
that the worst load on some of the POWER chips was more or less
nonsense code, designed only to thrash everything in sight.
Unless your job is testing CPU power management I wouldn't sweat
it trying to hit "100%".
about IBM's POWER5, but the figures were from Intel and presented at
IDF in 2004.
http://h21007.www2.hp.com/dspp/files/un ... heavys.pdfIf an activity factor of 1.0 represents the maximum device power that can be
extracted for any arbitrary code sequence (i.e. a power virus) then the maximum power of
any real application (typically Linpack) is around 0.8, SPECfp2k is about 0.7, SPECint2k
about 0.65, and TpmC about 0.6
So teamonkey even that paper you mention may not be an example of
Cell running close to "100%" i.e. Peak FLOPS isn't necessarily the greatest
workload.
Still I'm sure I've had a conversation about this, about 3-4 years
back arguing the toss about future CPUs for the Mac, and was told
that the worst load on some of the POWER chips was more or less
nonsense code, designed only to thrash everything in sight.
Unless your job is testing CPU power management I wouldn't sweat
it trying to hit "100%".
- teamonkey
- Sheepie Whore

- Posts: 12082
- Joined: Thu Jun 13, 2002 1:07 am 0
- XBL Gamertag: itsthetaste
- PSN: teamonkey_ca
- Contact:
Oh yeah, it was just an aside. The interesting thing about that particular experiment being that the proportion of calcualtions spent doing useful calculations in that paractical example were very close to the theoretical peak. It's almost as if Cell was designed with that sort of thing in mind.Kemper_Boyd wrote:Sure I've read that, did you notice all my comments?
- Kemper_Boyd
- Sheepie Whore

- Posts: 11967
- Joined: Wed Jul 31, 2002 9:15 pm 0
I was very sceptical about Cell in HPC, the programming, but I think it willteamonkey wrote:Oh yeah, it was just an aside. The interesting thing about that particular experiment being that the proportion of calcualtions spent doing useful calculations in that paractical example were very close to the theoretical peak. It's almost as if Cell was designed with that sort of thing in mind.Kemper_Boyd wrote:Sure I've read that, did you notice all my comments?
come to dominate the TOP500, the eDP Cell is going to be nerd catnip.
-
Kronos
Nah, I'm talking about utilisation of multi-processor multi-core type systems cpu's 100% in technical terms (ie all of the circuits on the silicon). Not what the PS3 or 360 are capable of as designed, but as 100&% utilisation of all areas of the chip at all times to do something more productive than running a 10 instruction independant loop on each for example. Moore's (I think) law or something about parallelism in processing.Snakey wrote:Old_but_still_a_kid wrote:Oh well, then the 360s peaked and M$ is stuffed in a couple of years time
So given that Motorstorm is throwing about 300% more at the PS3s GPU than it can physically render at 60fps, and it doesn't matter what they do with the rest of the system, it's still not going to be able to render that any quicker - I guess the PS3 peaked before it was released
As KB says you can overload any system, that doesn't mean it is being 100% utilised though in it's utopian maximum of 100% physical use, would need the perfect balance of independant processes and perfect fetch/execute timing of code to I guess. So to reply to Fdave's comment, you can certainly create code that max's out any cpu or gpu and leaves it wanting (look at all the shit games launched with 360 and compare them to what Gears of War is doing
Load of crap. Doesn't matter if you can process 500 billion polygons a second if you can't render the fucking things.Old_but_still_a_kid wrote:Nah, I'm talking about utilisation of multi-processor multi-core type systems cpu's 100% in technical terms (ie all of the circuits on the silicon). Not what the PS3 or 360 are capable of as designed, but as 100&% utilisation of all areas of the chip at all times to do something more productive than running a 10 instruction independant loop on each for example.
Yeah, because everybody wastes their time writing "bad/bullshit code" just for a laugh, don't they?Old_but_still_a_kid wrote:Anyone can write bullshit/bad code to kill a cpu, whether you can write a game that uses ALL parts of the system AND does it with 100% EFFICIENCY then and only then do you have 100% utilisation. That's the key.
It's really quite simple.
1. Write some code that does a hideous amount of very real and very useful processing.
2. Optimise the crap out of it.
3. Run the same code on multiple cores.
4. Watch as the performance goes up in exactly the way you expected it to.
5. Done.
The same thing is much more difficult on PS3. "Dr. Phil" seems to be saying nobody can do it. He's probably right.
Exactly. Cell is bloody good at this sort of thing, because that's what it was designed to do. But how much time is going to be spent doing this sort of thing in a game? Is a game ever going to get anywhere near peak performance out of it?teamonkey wrote:The interesting thing about that particular experiment being that the proportion of calcualtions spent doing useful calculations in that paractical example were very close to the theoretical peak. It's almost as if Cell was designed with that sort of thing in mind.
We need to forget what Cell is and remember what PS3 is.
Nothing is worth the effort you put into it.
Simple fact is you can't. GPU's in particular; there are certain places on the chip where you have to overengineer if you want to work well on a wide range of loads. So you might be able to triangle setup 500M primitives a second - because if you have a batch of 10-pixel triangles come through you need to do that or the pipeline bubbles to hell. But the majority of the time you're rendering 1000 pixel triangles and the faster setup rate matters nary a jot.Old_but_still_a_kid wrote:whether you can write a game that uses ALL parts of the system AND does it with 100% EFFICIENCY then and only then do you have 100% utilisation. That's the key.
Bottleneck analysis is more useful, but even then - well, performance analysis is bloody hard.
-
Kronos
Yes you can run the same code on 3 cores and it will run faster than 1 core, but will all three streams run independantly and never have any dependancy of the results of the other two streams/processes at any time. If you can do that then yes they will happily F/E independantly ad-infinitum. But doing this for the entire code all the time for a game has been done?
Of course it has.
Multiprocessing is far from a solved problem generally, but assuming you have enough memory a double buffered gamestate technique makes sense (and potentially solves some physics issues, such as the results changing depending on the order of evaluating objects). That would scale pretty much linearly with cores when well implemented.
That only leaves building and kicking off the actual command buffer(s) in serial code, and that could be a very simple operation, and from then on Amdahl's law gives the rough performance number.
So overall the basic principles aren't extraordinarily difficult. Implementation is far from trivial, mind - there's a lot of details about working sets, data arrangement and cache policies that needs to be got right even on a shared memory architecture, and I would be much more nervous about managing memory traffic on a Cell-type architecture where the gamestate object working set needs to be kept inside the size of the SPE buffer, and has to be pulled into there by scatter DMA.
Of course, this is a simple pipeline. One that uses the GPU for non-rendering tasks as well will be a lot more complicated.
Multiprocessing is far from a solved problem generally, but assuming you have enough memory a double buffered gamestate technique makes sense (and potentially solves some physics issues, such as the results changing depending on the order of evaluating objects). That would scale pretty much linearly with cores when well implemented.
That only leaves building and kicking off the actual command buffer(s) in serial code, and that could be a very simple operation, and from then on Amdahl's law gives the rough performance number.
So overall the basic principles aren't extraordinarily difficult. Implementation is far from trivial, mind - there's a lot of details about working sets, data arrangement and cache policies that needs to be got right even on a shared memory architecture, and I would be much more nervous about managing memory traffic on a Cell-type architecture where the gamestate object working set needs to be kept inside the size of the SPE buffer, and has to be pulled into there by scatter DMA.
Of course, this is a simple pipeline. One that uses the GPU for non-rendering tasks as well will be a lot more complicated.
Absolutely. No point doing it otherwise. Just get them all working on an equal but different part of the same data, so they all finish at the same time with the final product. Then you can put them to work on the next bit.Old_but_still_a_kid wrote:Yes you can run the same code on 3 cores and it will run faster than 1 core, but will all three streams run independantly and never have any dependancy of the results of the other two streams/processes at any time.
Incidentally, there are 6 hardware threads. Not 3.
Nothing is worth the effort you put into it.
This is quite significant - Virtua Fighter V coming to the 360:
http://www.computerandvideogames.com/ar ... ?id=152824
So the 360 gets a decent, the very best, fighting game, and Sony lose another shit hot exclusive. Interesting times.
http://www.computerandvideogames.com/ar ... ?id=152824
So the 360 gets a decent, the very best, fighting game, and Sony lose another shit hot exclusive. Interesting times.
I've seen people play VF4 and I was imressed. When I tried to play though all I managed to do was small punches.Rod_S wrote:This is quite significant - Virtua Fighter V coming to the 360:
http://www.computerandvideogames.com/ar ... ?id=152824
So the 360 gets a decent, the very best, fighting game, and Sony lose another shit hot exclusive. Interesting times.
Armed with a Cintiq+4+bless I'm off to fight bad taste!
Take a look at my art and t-shirts
Take a look at my art and t-shirts
Yeah it's not exactly accessible and pick up and play in the way Tekken and DOA are. The fighting system is deep, demanding, and requires a significant time investment if you want to get past the button mashing, flailing around stage of your first hour with it. If you can click with it though you'll see why it kicks the shit clean out of every other beat 'em up around, it's just so much more sophisticated, focussed and satisfying. I like Tekken and don't mind the occasional blast on DOA, but VF truly slaughters them in every regard.grue wrote:
I've seen people play VF4 and I was imressed. When I tried to play though all I managed to do was small punches.

