Everyone seems to like jumping off the handle on why PS3 is the fastest / bestest console ever. Now this is NOT a negative post about the PS3. Only a realistic post based on a certain amount of expertise (masters degree in computer engineering)

Here's why Cell is fast and efficient. When Intel makes the pentium, give their engineers 100 million transistors, and they'll make a processor that can do let's say 1GFLOP of performance. Give them 200 million transistors and they can do 1.5GFLOP of performance. Give them 400 million transistors and they can do 2GFLOPs of performance. These are made up numbers and inaccurate but they demonstrate my point. You get diminishing returns as you add transistors to a processor to speed it up.

You get diminishing returns because when you get more transistors you have to be more and more creative with how you improve performance: tweak the branch prediction, add cache, add support for more complex/specialized instructions like MMX. When you add a speedup to your processor, you pick the one that gets you the most bang for your transistors. This means when you go to choose your next speedup, you have to choose something less worthwhile.

CELL did it differently. They designed a scaled down processor, the SPE, where it has minimum transistor count and all RAW SPEED. Without getting into the diminishing returns on small speedup tweaks. Essentially this means that Performance per Transistor on the SPE dominates things like the pentium. Then to add performance to their chip, they simply added copies of the SPE. Why doesn't Intel do this? because computer programmers haven't been smart enough to write threaded, parallelized programs in general and they want a core that can run one thread as fast as possible.

That's why CELL is fast.

Now, the nVidia RSX is a mystery. We all don't want to think of it as just a 7800gtx, but too many signs point that way.

7800gtx - 300 million transistors
RSX - 300 million transistors

This is THE WORST news that we have so far. This means that whatever PS3 / CELL specific optimizations that the RSX contains, must come at the expense of some other transistors.

The big question is. In the 7800 GTX, are there many transistors that are used for tasks that computers require of the GPU, that the PS3 might not require because the CELL is such a single point floating precision monster? How is nVidia taking advantage of the fact that this is not a computer, it's a console. To finely tune the graphics for the PS3's exact needs.

We all hope this is true. Because including the 10 MB eDram the Xbox360 has even more transistors. And while I think that nVidia has been dominating ATI for a while on GPU performance, I still think ATI with 330 million transistors at their disposal can match nVidia's 300. But the question will be:

Does the eDram help ATI's performance a lot. 100 million transistors worth.
Is RSX specialized for the PS3

This SME thing is a load of crap IMHO. You LOSE performance if instead of having SPECIALIZED FOR GRAPHICS shaders you have GENERALIZED processors. This is a freakin' rule of nature. When you specialize in something, your flexibility is lowered, but you get better results at the narrower thing you are good at. PERIOD. This is why GPUs specialize at graphics and they will ALWAYS be better at graphics than CPUs. Don't fall for the Cell beats GPU at most, or all graphic tasks. The key is. it beats GPUs at a FEW graphic tasks that GPUs WEREN'T DESIGNED FOR (or say this another way, they weren't specialized for).

Really we all want the PS3 to be the best console it can be. CELL is already looking like an ideal chip for the PS3, minus the fact that the 8th SPE is unavailable.

We're all a little nervous because RSX is such a mystery and so much about the PS3's graphic performance will hinge on the RSX being good. But really if nVidia was to make an anouncement that we all wet our pants in excitement for how awesome it will be. It will be for one of the following reasons:

1. A change to the way software interacts with the GPU because it's a console and they can dictate massive changes (unlike computers where the APIs are set in stone) allowed them to rework the GPU in a drastic way increasing performance (this is the fundamental reason why CELL is such a beast. They announced that normal coding wouldn't work well on CELL because they required a new way of coding that would provide programs that would run on a brand new architecture that had huge advantages over old CPU techniques)

2. The fact that RSX is in a console, and paired with CELL allows them to not have to do certain tasks that were required before. This allows them to be reconfigured to do more productive work. (this announcement means that PS3 allows RSX to be even more specialized than a normal GPU already is specialized, for example if they announce that CELL will do ALL vertex transforms and vertex shading. Then they could include 0 vertex pipelines and use more pixel shader pipes. This is MORE specialization. See?)

3. They've made it more expensive. (Increased transitor count, or 2 RSX in SLI)

nVidia and ATI have SMART employees. There is no LOW hanging fruit. What this means is, if there was some AND gate they could include in their chip that would make it 1000000000x faster. They would have done it. The existence of some simple, easy, no reason not to do it HUGE performance increase that has NOT been implimented IMPLIES that ATI and nVidia are stupid. Really. if there was some no-brainer choice that would make their product rule supreme and they had just missed it. I don't think so. I'm sorry, they have hired too many great minds.

To summarize:

Are GPUs the fastest they will ever be? No of course not. But true power improvements come from high hanging fruit. Stuff that is harder to get to, or stuff that there are reasons not to do it. Where you have to sacrifice something else. And all I'm saying is that the three main ways to get these things are to:

1. Change radically in a way that requires the users of your product to change the way they work with it.

2. Become more specialized.

3. Add cost (second chip in SLI. bigger die size. dry ice cooling and overclocking)

It would add back transistros being taken out by unusable PC functions. Perhaps we are taking the transistor count too literally when comparing both. There are one story houses that have as many square feet as some two story houses and vice versa. Again, I don't think the Nvidia gurus were sitting around drinking coffee on Sonys dime just expecting to solder a 7800 core onto a PS3s system board and call it a day. Sony is doing the manufacturing themselves. Nvidia had absolutely nothing to lose by putting the G70 back on the drawing board for the PS3.

Yes, when you for instance, (Play Along), press cntrl alt delete, go into the processes tab.

Now, for the simple process of going onto the internet (which is ALL the ps3 would have to do) it uses... 39,136k
And, that... is no where NEAR 1/4 of my whole usage.

Ps3 will be able to centralise all power on the process. Whereas the pc will still be running 50 programs you don't need.

Even if it was the 78000gtx .... its still gonna be ****ing good at it =]

And i doubt it is..

While there is extra transistors for those things, I do believe that they are not completely separate units. Most of the processing power to do such work is already in the GPU, being used for other functions.

Also, adding the equivalent of turbocache onto the GPU would ADD transistors.

Don't forget that Cell has a PPC unit as well with 512 cache. That's plenty enough to run a modicum of general purpose tasks such as an email client or a browser. It's certainly not an AMD FX60, but it will suffice, especially for the needs of what the PS3 is designed for.

FLOPS is short for floating point operations per second.

Why can Intel chips do so many fewer FLOPS? Well, most of their circuit logic is meant for things OTHER THAN floating point operations.

The cell was designed to be a floating point monster, but is actually not very powerful at other operations.

It is also highly optimised for streaming data applications, and code run in parallel.

So take note, this is a specialized approach.. I and others might argue its somewhat too specialized... It can easily be bottlenecked by being overwhelmed by non-floating point operations.

Games are heavier in floating point code than say internet explore or say microsoft outlook. Classically, games have not been proportionately floating point in the way cell has organized. However, it remains to be seen whether or not programmers can change this and utilize the FP power of cell. If they can use those SPU's, it will blow the competition away.

As far as the RSX, I think its quite likely that its essentially a G7x running at 550mhz. I don't think this is bad news... It will still technically be more powerful than Xenos, but it just won't be by a very big margin...

I'm of the opinion that they took out a lot of the fluff like TrueVideo and other stuff that PC GPUs need. That die-space (transistors proper) are being used for something else.

Like you said, these people are smart, and they are probably hinging a lot of performance gains from the very fact that Cell can work with the RSX via the Flex I/O bus. I know the good folks at Nvidia weren't sitting around with thumbs up their butts staring at a G70 core and saying "that's it" for this long.

But Sony is also assuming that developers can learn how to work with multiple threads, a task that a lot of PC developers have yet to do.

Very coherent. Sony has had the chance to design a powerfull chip from zero, contrary to intel/amd who are forced to keep retro-compatibility with every x86 processor. This has allowed them to really think of what kind of calculations a next gen game would need, and intel can't do that.

chip info 101? *takes notes* :P

Yes you are correct... I was merely illustrating that the designs are meant for different implementations.. Cell was aiming exclusively for FLOPS, while your average Intel processor isn't really designed for FLOPS... Take the server market... There are often a lot of processes going on in your average webserver... You aren't using quite as many FLOPS here..

PowerX architecture has always been good at multimedia creation and such... This is due to the Altivec/VMX portion of the CPU. Of course, you wouldn't use this for distributed computing or anything like that...

Cell will make great scientific workstations, encoder/decoders, rendering stations, etc etc... There are just things that it won't be good at... which is why you might not see it into a PC until it adds and improves its PPE's.

I think this is why Sony/Toshiba/IBM are continuing to work on this... They will probably try and make the design more plausible for every day use.

Woe be the day when STI get the Cell down below a 35nm process and start adding more LSRam to the SPEs and cache to the PPE... Cell is already a generational leap beyond multicore processors as it is, but dropping below 35nm is going to be where the real bread and butter is at.

You really know your stuff, man.

will this is great info. cell and RSX was built for this. they are going to make sure that ps3 live up to the hype of power.

I wasn't knocking Intel in my original post. Only pointing out that they are less specialized than CELL (this is a neutral statement. Making Intel better for some things and Cell better for others).

I think in general the parallelization of programs isn't too specific. There are hardly any cases where you can't thread some kind of problem. Even stupid programs like Word and Excel rely on algorithms, data structures, and shared libraries that have functions that could be ported to run simultaneously on SPEs.

I can't really think about much in computing that both requires a lot of processing power AND can't be parallelized on something like an SPE.

In fact the only thing that I can think of right at this very moment is: Emulation. Simulating another chip. It seems like for this, the only thing you would want is high clock speeds, branch prediction and no cache misses.

But you have to see that Cell is designed to run media and gaming applications well. And also, general computing in general will be no slouch on it as well because it is really forward thinking. Why do you think Intel and AMD are both rolling out multi-core chips, with plans to make them even more multi-core than dual-core in the future? Cell is ahead of its time. Which is good for a console that is supposed to be around for 6 years.

Well to answear your question about RSX, yes the PC 7800 does conatin alot of ******** that RSX does'nt need like

1. Nvidia pure video
2. Video encoders
3. DX9 logic, i know DX9 is software but GPU's do cantain a very SMALL amount of logic to help DX9 along
4. Cine FX 4.0
5. The SLI logic.

Plus loads more Stuff...

There is loads of crap that it does'nt need, the question should really be..

What have they used those spare transistors on???

And IMO, the EDRAM was a watse of tranny's

I agree with antuk15.

The only reason I care is that there are too many threads that go: Such a such a board had a post from a user saying nVidia has incorporated MagicPixieDust(tm) into the RSX and now it is reported to be 10,000 times faster than 10 of the worlds fastest video cards chained together.

I don't think we shouldn't speculate about how the RSX is going to be strong or weak. I just think when you speculate, you should speculate about how the RSX is going to be worked in a more realistic way. Such as, it's implemented something in place of antuk15's 1-5 list with a new CellLink circuit which does X Y Z. or whatever.

THe x86 or intel processor is a great design but one of the major things that holds its development back is it has to be compatible with everything. The reason why the processor has been so successfull is the fact that it can pretty much run everything for the past 20 years, its design has been improve but they can't make major logic changes to the processor because if they did it would force people into updating very expensive legacy applications.

Cell allows a clean slate, its always touted as to why PowerPC & other chips are more powerful then Intel but less used, its mainly the compatibility.

Sh?t...I don't understand all this computer uber speak...

Simple english please...

St Jude. The translation is:

The more you break off convention and not follow the past. If you give up being backwards compatible with previous technology, you can invent something that is SPECIFICALLY designed for the task at hand. Doing this you can create something that is revolutionary instead of evolutionary in terms of power.

CELL has done this.

The big question is has nVidia done this with RSX, or have they stuck with convention and made an incremental improvement.

I would say they would have had to break convention.

A lot of the processing would probably be similar to current but the major change they would have had to do is make it work directly with the cell,

They have designed a system where the CPU & GPU can work in tandem & work directly with each other, you can't just rip out a current gen PC GPU & make it do that, you can use some of the well developed technology, but they still would have had to do it from the ground up.

The EDRAM was certainly not a waste of transistors in the case of the XBOX. Without it, the system would have not been able to do much.

The shared architecture for the XBOX 360, along with the measly 22 GB/s bandwidth could be topped by something along the lines of a Radeon 9600. The EDRAM is VERY fast, and saves a ton of bandwidth by handling the color + z data, plus most FSAA and HDR implementation. 22 GB/s would not have been enough otherwise, and being shared certainly does not help either..

You will never see it on a graphics card though.... It would be cheaper to put the extra bandwidth on the shoulders of graphics memory...

So, EDRAM was a great idea for a console after the cost cutting maneuvers they pulled on it.

nick? makes some good points.

We all like to talk about the 360, but the truth is microsoft has been kind of tight lipped about the technical specs and capabilities of their console really. So there is still a ton of speculation really. People who don't know what they are talking about look at the huge EDRAM bandwidth and go, look how much more bandwidth the 360 has than the PS3. But really the EDRAM isn't extra bandwidth, it's offloading the bandwidth that otherwise would have competed for vram access, so to really know how much the EDRAM is contributing to bandwidth, one has to ask, well how much would have to go over the main bus if the EDRAM wasn't there, and this is a good question. So it's not useless, but it's hard to say how much use 360 is getting out of it.

We all like to think that because the 360 is out, the cat is completely out of the bag, but it's still pretty much a MS secret. Just because they've released some PR clips with a few bits of information, you know they only release things that they feel will make the console look strong, not enough to give you the whole picture.

I agree.. most people don't seem to think this would take a lot of time to do.. I think that considering the cell is a new technology, this would have taken a considerable amount of time to integrate and do right.

I think the G70 core will still have to lay at the heart of it all... It wouldn't make any sense otherwise... Other than adding more pipes, there is no sense in separate shading devices....

Of course, they could have always added on something like EDRAM... because I don't see how they are going to compensate for lack of memory bandwidth... which is what I see as the fatal flaw in this setup...

I would imagine anything added would have to alleviate these issues, as with the XBOX..

Yes, its merely a way of giving the Xenos core the same abilities as if it was on a standard graphics card. Its nothing special... The AA and HDR really isn't free so to speak... Its just that it isn't taking the remainder of the (shared) 22 GB/s meant for mostly texture bandwidth.

Cell is evolutionary, its coming from designs from years ago, however the increase in the clock speeds at the time outweighed the benefits of going this route... however with the "clock barrier" we have essentially hit we are starting to need to do a lot more to work around memory latencies and clock speeds. Ideally what we want is a really really fast processor (no logic required to share data or sync the caches) with a 1 cycle (or even a small number) memory latency... but as we get faster the latencies get bigger and we can't go faster without melting chips.

RSX and suchlike are basically simple evolutions each time, occasionally we get a big jump but like the P4 they are essentially backwards compatible back to the stone age of technology. Consoles are great for this though as we can afford to be compatible with virtually nothing and very few people care!

Itanium was supposed to be Intel's way of getting off the x86 technology... It flopped, mainly due to no software support...

Oddly enough, Cell and Itanium share many of the same concepts... Itanium features an 8 way parallel design, and works very good as a DSP... but not as a general purpose processor...

Development of the platform really hasn't taken off... and will probably die out (save special circumstances), much like MIPS.

There is saftey with the x86, all large projects look for maintainability, no one is going to spend several million dollars on a system & not worry about the future, software can last 20 years if its implemented well, hardware needs to be upgraded.

the x86 family provides the certinty that you wont have to rewrite your software in 3-5 years when you upgrade the hardware.

IT projects get expensive & most project managers look at the most cost effective ways of maintaining their systems.

Yes, oddly enough some users at my company still run DOS programs... The simple fact is, new software + hardware costs on top of user training... it can make it not worth it to upgrade. If the users have to relearn their tasks, just to use new software, tons of problems take place... If they start doing their job wrong because of the software, who's head will roll? Probably the IT managers...

I think there is a good deal of emulation avaialable, but this always runs slower than the native platform.

I love it when the naysayers gang up on the Montecito (Itanium's next iteration due 2006). What many people don't realise is that the Montecito has time on it's side. CPU architecture is not going to stay with x86 for much longer. You can see this already. The CELL despite all this "revolutionary" talk is no more then an advanced Itanium architecture based on a Power core rather then the Intel core design.

Here are some pieces from an Anandtech article that I like, since it sheds some light over the whole Itanium fiasco:

Main advantages of the Itanium:

Do those look familiar to anyone?

Now here is a bigger piece of that article that looks verry, verry interesting:

Anyway, here is the Link to the whole thing.

Now, I've been interested in the Itanium ever since it was first announced (and have stuck by it even with the Merced deal) and I do believe that this will be the way Intel wil redeem themselves in the CPU industry... just wait and see.

Thanks for that.

The itanium is a really good processor, it was just really brought out ahead of its time.

Now with cheaper ram & much better emulation, it is easier to start moving platforms & providing the same backwards compatibility, the next few years are going to be good to watch.

I remember in about 95, 96 RISC CPU's were being marketed then as "the next big thing"

they took off in handhelds & portables but the PC market just kept growing, the transition away from x86 over the next 5-10 years is going to be a pain in the *** though, like all migrations there are always problems.

Well, we wil see.. If the whole market shifts in the direction towards cell like architecture, and away from x86, it is likely that it could be revived...

However, any non x86 architecture better be very tempting... or its not going to pull it off.

Even though I am not technologically minded, what you guys are trying to say is a Cell Proceesor can kill a Pentium 4.

in gaming ? yes
in floating point calculations ? yes
in streaming multiple hd movie streams ? yes

in doing the usual ms office business stuff ? no

people will stick to well known pc`s even when cell is out. i wonder if we will see a cell workstation at all.

To answer that question it's a YES with an IF or a NO with a BUT

NO it can't beat the pentium for general computing applications (as they are currently written now) BUT if developers all switched to a CELL like developent method it could

Yes the CELL, and CELL broadband architecture can kill a Pentium 4 IF all applications were rewritten for CELL.

People don't just wake up and decide to rewrite 20 million dollars worth of development spent on their current software projects unless they have the idea that it'll get them more than 20 million in value.

This is why you don't see people jumping at the prospect of completely re-writing MS Excel so that it can run 40% faster on CELL than it does on the current P4. Because nobody needs a 40% faster Excel, it wouldn't generate any extra revenue.

So you have media/gaming/stream processing/scientific calculation where CELL really exceeds switching over because it makes sense to put in all that rewriting of code to gain in performance.

I see...

But why can't the Cell do both...why does it have to do only one?

Yes but what most people dont understand is that its not "FREE" the geomatry overhead it causes 360 is massive.