Was Cell any good? *spawn

[Arwin]

It happens in more games. In Uncharted for instance when you walk around in areas with a lot of NPCs, some will be looking at you, following you with their head movements and eyes.

EDIT: like here, when they walk through the market (don't watch if you don't want some minor spoilers for this game). You can see it much better of course when you're just playing and not rushing trough like guys for these complete walkthroughs always tend to do.

 

[Shifty Geezer]

I am a weakling at technical matters, but I am following the discussion and I wonder if what happens in this video can be considered ray-casts:

Could well be, tracing a line of sight from NPC to player.

 

[sebbbi]

That's a great example of raycasts

Basically when you do something, the game checks if any of the NPCs had a direct unoccluded line (ray) from their head to your character. If the raycast hits something in the way, then the NPC cannot see you. Naturally if you want to make that scenario more realistic, the NPC should react in some way when you drop the pot on it's head... or at least it could also react to the sounds around it. It's a really funny video nevertheless

 

[upnorthsox]

I am a weakling at technical matters, but I am following the discussion and I wonder if what happens in this video can be considered ray-casts:


I always thought that any character in a videogame don't have eyes, as a sense, that their eyes were just there for physical accuracy and depict natural proportions of living beings.

I mean, in a game, theoretically, they should be just a body, and "their eyes" could be just all over their body without making a difference, since you can't picture what non playable characters and creatures see no one would bother to simulate the sense of sight. Developers could only draw a person correctly and get away with it, because it seems enough to treat a body as a whole, without distinguishing between the eyes and the rest of the body. In this case it's obvious an object placed at eyes' height is occluding the non playable characters vision, and it looks realist.

This seems like you posted in the wrong thread...............but anyhow, in that shot it'd probably be more accurate to say that line of site is located around the head area, not necessarily from the eyes. That completely ignores the fact that bartender doesn't seem to care that you are about to put a kettle over his head. His whole reaction could be based off the items being added to your loot list.

 

[patsu]

As I understand, one of the most difficult tests for Cell is H.264 Main/High Profile's CABAC decoding. Data access sounds pretty random/dynamic because the operation changes depending on the data. For every bit/pixel, the algorithm needs to choose different probability model to use, and at the same time use neighboring elements to optimize/update the model on the fly (for processing subsequent bits). So the process sounds somewhat serialized too.

Before BR was released, some people claimed the algorithm is impossible to implement efficiently on Cell (1080p @ a throughput of 40mbps). Anyone know how they did it ? Do they use the PPU more in this case ? They did it with resources to spare for a Java VM and a second 1080p stream.

EDIT: Or do they use the SPUs to perform speculative work in parallel, and/or throw more cores at the problem to bruteforce through the slow process ? In which case, the PPU may be used for running Java.

 

[patsu]

While looking into Cell BR run-time, I found this paper showing how to run JavaVM on the PPU and SPUs:
http://www.lst.ethz.ch/research/publications/CF_2008/CF_2008.pdf

The heterogeneous design of the Cell BE architecture provides a significant challenge in supporting languages such as Java and C# which are implemented with a virtual machine.
...

CellVM [11] is a port of JamVM to the Cell BE. It consists of two distinct JVMs each running on separate cores. ShellVM runs on the main core and distributes work to each of the SPE cores on the Cell BE. A second JVM, CoreVM, runs on an SPE and executes the Java bytecode supplied by the ShellVM. A limited number of instructions in the JVM instruction set do not run on CoreVM. These include instructions that create new objects (including arrays) that must access the heap stored in main memory. A switching mechanism is employed to revert back to ShellVM processing for these instructions

Perhaps the PS3 Playstation Suite C# run-time will use something similar.


The VMX is only a vector math engine. So it's not generally tasked to run a JIT VM (fetch, compile and execute bytecodes).

EDIT:

8 Conclusions

This paper has presented analysis of existing interpreter optimization techniques on the Cell BE Processor and introduced novel optimizations made possible by the architectural features of the Cell BE SPE. Firstly, interpreter pipeline techniques first published by Hoogerbrugge et al. and tested on the Trimedia VLIW processor were tested and shown to provide increases in performance on the Cell BE. Further, these techniques were extended by introducing an optimization referred to as operand pipelining in which the existing pipeline structure was ex- tended by performing a decode stage in which operands were fetched in advance. This technique proved extremely successful providing a speed increase of just under 20% on our best performing benchmark. Secondly, a novel optimization for the dispatch of conditional JVM instructions was described. This technique used the hint feature of the Cell BE to improve the dispatch performance of JVM instructions that compute their control flow target. Finally, the limitations of the Cell BE’s abilities to interpret JVM bytecodes efficiently were identified. They include the large hint to branch stall (approximately 15 cycles) and the penalty incurred by expensive load and stores (6 cycles). The opportunities of the first limitation were demonstrated through a useful technique that eliminated the number of JVM instruction implementations required by performing a merging and pairing optimization co-locating similar instructions in the same implementation. These merged instructions were subsequently used to provide a rich environ- ment for producing superinstructions combinations that produced an increase in speed over many benchmarks. Through the use of software pipeling an efficient pipelined interpreter was designed that removed the load and store limitation of program counter fetch and decode as well as prefetching operands.

 

[liolio]

I don't want to spam the WiiU thread which is (supposedly) dedicated to speculation on the GPU (may be should be renamed?) so I'm posting here.

I definitely believe that the Cell was indeed good though oddly balanced too many SPU and sucky CPU.

I do believe that within what looks to be pretty severe power and silicon constrain for their next system, the WiiU, Nintendo should have a use a Cell2.

Something like a power7 and 3/4 SPUs. I believe that it would have allowed the system to keep up with the PC realm for the life time of the system and that it was the cheapest way (in silicon and power consumption) to do so.

The SPUs are impressively power efficient and consume next to nothing (especially if they were to run at a lower clock speed) and they are really tiny, tinier than mobile CPU once you take in account the L2.

Overall I could have seen a "slim" Power7 passing on Edram (the L3 would be made of sram and so faster but in lesser quantity), and the SPU would be connected to a crossbar instead of a bus.
I would expect the thing to be above 100 sq.mm on IBM std 45nm process.

So I'm not asking my-self if the Cell was any good but if it could have been beaten within the constrain Nintendo seems to have for its system.

 

[Prophecy2k]

I don't want to spam the WiiU thread which is (supposedly) dedicated to speculation on the GPU (may be should be renamed?) so I'm posting here.

I definitely believe that the Cell was indeed good though oddly balanced too many SPU and sucky CPU.

I do believe that within what looks to be pretty severe power and silicon constrain for their next system, the WiiU, Nintendo should have a use a Cell2.

Something like a power7 and 3/4 SPUs. I believe that it would have allowed the system to keep up with the PC realm for the life time of the system and that it was the cheapest way (in silicon and power consumption) to do so.

The SPUs are impressively power efficient and consume next to nothing (especially if they were to run at a lower clock speed) and they are really tiny, tinier than mobile CPU once you take in account the L2.

Overall I could have seen a "slim" Power7 passing on Edram (the L3 would be made of sram and so faster but in lesser quantity), and the SPU would be connected to a crossbar instead of a bus.
I would expect the thing to be above 100 sq.mm on IBM std 45nm process.

So I'm not asking my-self if the Cell was any good but if it could have been beaten within the constrain Nintendo seems to have for its system.

Surely the fact that Sony owns the design would have made it a non-starter?

 

[Nano]

Well, to me its clear the Cell processor is its biggest asset yet its biggest disadvantage.

Almost disastrous direction and definitely hit Sony financially as well as teaching them a hard lesson (taught IBM and Toshiba lessons too though they probably didn't suffer not having a major consumer electronics product).

However, it has provided the necessary computing power for good looking and well performing games. There are a lot of technical limitations a lot of hoops to jump through, but the overall result is reasonable and in cases really good.

Next generation Cell simply isn't a viable option.

 

[Nano]

Without the cell, games like GT5 would have looked like an absolute piece of crap due to the weak RSX.

Why do people take such a harsh line?

The RSX isn't that incapable that it can't produce good graphics. You wouldn't have the extra power from Cell but I doubt games would just look crap.

The RSX has the shader power of the Geforce 7800GTX 512MB though it has the lower bandwidth. Developers would have made games in a different way based on the strengths and limitations of the architecture.

If there was no Cell processor, well, the entire architecture of the console may have been completely different...

 

[liolio]

Surely the fact that Sony owns the design would have made it a non-starter?

Sony doesn't own the design, any of the three companies (Toshiba, Sony and IBM) can do whatever they want with the design.
IBM came with the ddr2 version and Toshiba with a SPU based design only.

 

[Nesh]

Why do people take such a harsh line?

The RSX isn't that incapable that it can't produce good graphics. You wouldn't have the extra power from Cell but I doubt games would just look crap.

The RSX has the shader power of the Geforce 7800GTX 512MB though it has the lower bandwidth. Developers would have made games in a different way based on the strengths and limitations of the architecture.

If there was no Cell processor, well, the entire architecture of the console may have been completely different...

Well the RSX can do good graphics, but with everything else given, the Cell is what helped the overall design reach that level of quality that made the PS3 competent compared to the 360

 

[liolio]

Well the RSX can do good graphics, but with everything else given, the Cell is what helped the overall design reach that level of quality that made the PS3 competent compared to the 360

Indeed still from a business pov I can't help but think that IBM was the main winner in the STI alliance.
Cell did the job inthe ps3 but it costed Sony quiet some money, if the design had some success the only actor out of the three that could have push it further were IBM.
Simply Sony and toshiba can use any part of the IP but they can't come with a new PPC.
Once IBM won over Toshiba idea to use many MIPS cores and the design was locked with a ppc as the brain looking forward they were the only one that could push the design further in any significant way.
For example if it were successful (and profitable enough) IBM could have replace the CPU by a more beefy one instead of having plenty of opterons+Cell boards designs in the HPC world.
I guess it was not worse their money/efforts, from insights read here it's not a high margin market.

What could SOny or Toshiba have done? Actually the point stands with SPU I don't believe that either Sony or Toshiba have what it takes to come with improved versions, IBM can/could have.

Pretty much that deal is imho not good for Sony Toshiba, they took part in a huge R&D effort whereas looking forward they were not in a situation to rip of the benefits if the concept had really take off. IBM was the winner of the deal (along with Rambus, which make money on something else than patent troll for once ).

 

[aaronspink]

Well the RSX can do good graphics, but with everything else given, the Cell is what helped the overall design reach that level of quality that made the PS3 competent compared to the 360

But without the boondoggle of Cell the whole console would of been better designed and they wouldn't of had to beg and plead at the last moment for a GPU.

 

[TheWretched]

Yes but imagine if Microsoft would'be been out of luck with their GPU (i.e. ATI not being ready with their unified shader architecture or whatever) and then both, PS3 and 360, having a very similar GPU? Then Sonys advantage in terms of the CPU would've been the factor that could've completely changed the playing field.

Or the other way around, Nvidia having a faster GPU in the back of the store... it could've easily gone either way for Sony or MS.

 

[Nesh]

But without the boondoggle of Cell the whole console would of been better designed and they wouldn't of had to beg and plead at the last moment for a GPU.

That doesnt say anything about how good Cell is and it doesnt mean a better GPU couldnt accommodate with Cell's architecture if Sony knew that the initial idea of having only Cells to render graphics or an exotic GPU wasnt the route to go. Its Sony's change of plans and various business decisions that have let to the last moment weaker GPU. If RSX was closer to the 360's GPU in performance, we would have seen the PS3 having the edge. And that would have been the PS3 if Sony saw the design from the beginning as Cell+a custom up to date GPU from Nvidia or ATI.

 

[Shifty Geezer]

I guess the argument to that is whether Cell+Xenos would have been worth the cost. If PS3 had the same GPU as XB360, would the investment in Cell have delivered worthwhile benefits, or would it have been a waste of money where a simpler CPU would have done the job nicely?

 

[fearsomepirate]

Why do people take such a harsh line?

The RSX isn't that incapable that it can't produce good graphics. You wouldn't have the extra power from Cell but I doubt games would just look crap.

Remember, you're dealing with console gamers, who amplify marginal differences between pieces of hardware or exclusive software in order to fan the flames of forum wars.

Wasn't the RSX originally specced a bit higher, but had to be cut down due to low yields?

 

[BRiT]

Wasn't the RSX originally specced a bit higher, but had to be cut down due to low yields?

That's what I recall as well, though it could have been due to TDP issues instead of yields. The specifics escape me.

 

[Acert93]

iirc it dropped from 550MHz down to 500MHz. Too lazy to pull a reference...