PS3's Cell CPU is by far stronger than new Intel CPUs

erek

erek

[H]F Junkie
Joined
Dec 19, 2005
Messages
10,900
Pretty strong opinion post? Cell processor seemed extraordinarily difficult to optimize for.

"The Cell processor was (and apparently still is) a potent workhorse that could push some serious performance, but tapping its power was a convoluted, complex, and often frustrating, time-consuming process. Developers had to specifically optimize their games for the PS3 to squeeze out the Cell CPU's full potential, which few games actually did. "The PlayStation 3 had such a bucketload of power," van der Leeuw continued. "Making use of it and really getting performance out of the PS3 was hard, because you had all the SPUs and the power was not easy to unlock. You had to write a lot of special-case code. Once we were done and we got all the physics stuff and the ragdolls and there was so much stuff we did, it was good, but everything took a long time. And you need quite a skilled team to do it.""

Read more: https://www.tweaktown.com/news/6916...ll-cpu-far-stronger-new-intel-cpus/index.html
 
That article is total bullshit.
The purposefully cut-back IBM Cell PPC CPU in the PS3 was capable of around 150-175 GFLOPS (FP32), fully optimized, and an Intel 6700K x86-64 @ 4GHz (pre-numerous security patches) is capable of 512 GFLOPS (FP32), and that is a CPU from 2015.

I know this because I actively use a 4-node PS3 beowulf Linux cluster and know exactly how optimized and unoptimized code and programs run on those units.
While the IBM Cell CPUs were extremely powerful for 2006, they are paltry for 2019, and this article is beyond ignorant.

Also, the Cell was not meant to be used as a general-purpose CPU, and really was more of a stop-gap between how CPUs and modern GPUs operate, being heavily FPU-based (7-8 SPEs) with limited integer functionality (1 PPE).

All variants of the Cell were also officially EOL in 2012 by IBM, and for good reason.
Newer Power ISA CPUs were much more efficient, and even a mid-range Ivy Bridge quad-core had comparable performance with a much wider code-base (x86-64 vs PPC), and a much lower power requirement and heat exchange.


EDIT: erek, these comments aren't directed at you, and you always make great posts about really good articles. (y)
 
Last edited:
If this article was published 10-12 years ago it would be accurate. A Q9650 @4Ghz is good for about 46 Gflops so that makes the Cell look good. But something like a 3800X does 740+ Gflops, so no comparison there.
 
Also, FLOPS isn't everything.
Integer performance is still extremely important, and the Cell is abysmal at integer performance; this is understandable as it wasn't what the Cell was designed for, but that article was garbage...
 
The Cell in the PS3 wasn't just a CPU... it could be argued it was in a way still faster then what we are running now CPU wise. But a PC isn't just a CPU its a GPU as well. The Cell could be made to do the same type of math GPUs destroy CPUs at calculating.

The Cell was ahead of its time and is underrated greatly... this developer isn't wrong though it was unlike anything else around and optimizing anything for it was not easy. I'm not sure I would go so far as to say it was superior to a current i7/i9 or Ryzen chip. Again though it really depends what your talking about... if your talking about doing Physics calculations better done by a GPU... I have no doubt the Cell would walk all over any current CPU. So I don't think this guy is wrong in some very specific calculations ya the Cell is better then any x86 or general compute arch like Arm or Power... that really doesn't of course make it the better chip, as in most cases general compute chips are still better at general computation. :) lol

We all think Playstation with Cell... but it wasn't designed to just be a game console chip. IBM/Sony/Toshiba built a lot of power for vector processing and media acceleration. It really wasn't a standard CPU... the general purpose power bits where low power.

Anyway the article seems like clickbait but there is some truth to it. The cell chips are long gone now... but not that long gone. Roadrunner was shut down in 2013 and for period of almost a year till the end of 2009 its 13 thousand Cell 8i chips made it the fastest super computer in the world. They shut it down in 2013 cause it was just not power efficient enough to justify continued operation. Toshiba also used the cell in some high end TVs... I believe they could display something insane like 40-50 MPEG streams at the same time.
 
  • Like
Reactions: Mega6
like this
People fail to recognise the parts that make up a modern cpu.

Nearly every modern intel processor has a GPU on board. It’s not a screaming 2080ti, but it would put most (if not all) of the pre 2010 cards to shame.

Add to that 4/6/8 fast cores with AVX, and there is no way a cell chip of 13 years ago is faster. There is just no way. It’s not even faster by a little bit.
 
Keljian said:
People fail to recognise the parts that make up a modern cpu.

Nearly every modern intel processor has a GPU on board. It’s not a screaming 2080ti, but it would put most (if not all) of the pre 2010 cards to shame.

Add to that 4/6/8 fast cores with AVX, and there is no way a cell chip of 13 years ago is faster. There is just no way. It’s not even faster by a little bit.
Click to expand...

Accept it is for the most part... the PS3 Cell is not however a general purpose CPU, so its not a fair fight.

You could say basically every GPU for the last 10 years is also faster then any current CPU if your taking about doing FP8 vector math calculations. The PS3 was a hybrid... it a had a Power based PPE, it was very close in design to the of the day 64 bit power chips, although it wasn't. (IBM used a version of that core for the xbox) On the PS3 Cell the PPE controlled 6 SPE (synergistic processing units) each SPE had 6 cores with its own embedded L1 cache system... there where some serious issues programming for them as they couldn't directly access system memory among other PITA issues. However in a PS3 cell operating at 3.2 GHZ each unit had a theoretical output of 25.6 Gflops of power. Of course this isn't general compute math... and the Cell also lacked things like branch prediction. The developer in this article is actually spot on the PS3 IF (and that is a big if) is programmed by someone that knows how to setup proper virtual memory maps for all of the SPE units... and they have optimized the hell out of their compiled code to make the lack of branch prediction a non issue you could in theory run software on the PS3 many times faster then MOST modern CPUs, and no doubt many times faster then anything on the market during the PS3 days.

The PS3s combined cell hardware operated at a theoretical max of 230 gflops. For reference not that long ago Intel advertised the Core i9-7980XE as their first teraflop CPU. And the 10980 is basically the same chip. So by that logic if a 18 core Intel chip is only theoretically capable of 4 times the calculation power of a PS3, well 4 and even 8 core versions are at best =. (again we are not talking general compute on a PS3... but it was a very innovative design it was just always better suited to super computer work then game consoles with all the compiler level tricks required to put even a fraction of that power to work in a game) For very purpose built bits of math though Cell arch was unmatched... which is why roadrunner was the fastest super computer in the world till the end of 2009, and by comparison that machine didn't even run that many CPUs. Modern super computers are using 3-4x as many chips in their arrays these days. The only reason IBM didn't continue developing newer Cell based chips was the rise of the GPUs. They basically do the same job.
 
Last edited:
Its funny we are still hearing this crap about Cell 13 years later. Well, let me know when Ps3 games that take advantage of this vast power are released. Maybe the Geforce FX5800 is just an optimized driver away from outperforming a 2080 too.

I'm a Playstation guy myself, but I can't really remember a single cross platform game that looked or played better on the ps3 vs 360. Was routinely disappointed reading review after review of games that came to this conclusion!
 
Insomniator said:
Its funny we are still hearing this crap about Cell 13 years later. Well, let me know when Ps3 games that take advantage of this vast power are released. Maybe the Geforce FX5800 is just an optimized driver away from outperforming a 2080 too.

I'm a Playstation guy myself, but I can't really remember a single cross platform game that looked or played better on the ps3 vs 360. Was routinely disappointed reading review after review of games that came to this conclusion!
Click to expand...

Hardware is only half the equation. Software was always the issue with the PS3. As the guy in this article says. The power was there in the hardware it was just extremely hard to unlock. Considering games at the time started using PC engines... its not shocking optimizing them for a completely foreign ISA was not just difficult but damn near impossible. Before the PS3 game studios mostly made their own engines and actually had engine designers on staff. It was around this time (and probably a bit before to be honest) that most game companies stopped hiring those types of programmers. Very few game studios have engine talent on staff anymore. (or they hire a wave of them every few years on contract to work on a new engine they then use for multiple games)

So ya the PS3s issues where basically 100% software related. It was a beast to program for and game studios with engine talent where focused on the more standard platforms... and more and more studios where starting to not have in house engine talent at all. Making optimizing for PS3 damn near impossible.

Sure the xbox of the day had a powerpc core... but it was a standard triple core power core that did general computation much like x86. Compiling for that machine was not that hard it was still general compute hardware and MS libraries basically made it a easy recompile.

The upside for Sony was they developed the core of their shader APIs back then... and it forced Sony to add a lot of quality of life things for programmers to their APIs. A lot of game designers prefer sonys APIs over all others today. No doubt the PS3 software development has helped Sony make the PS4 a system developers have been able to optimize very well.
 
As far as the mathematical calculations that the Cell Processor did for what it was designed to do, yes, even modern day processors would have a hard time keeping up, and vice versa. Could it run MS office ect. as fast as a modern day CPU? No, but that's not what it was designed to do.
 
ChadD said:
Hardware is only half the equation. Software was always the issue with the PS3. As the guy in this article says. The power was there in the hardware it was just extremely hard to unlock. Considering games at the time started using PC engines... its not shocking optimizing them for a completely foreign ISA was not just difficult but damn near impossible. Before the PS3 game studios mostly made their own engines and actually had engine designers on staff. It was around this time (and probably a bit before to be honest) that most game companies stopped hiring those types of programmers. Very few game studios have engine talent on staff anymore. (or they hire a wave of them every few years on contract to work on a new engine they then use for multiple games)

So ya the PS3s issues where basically 100% software related. It was a beast to program for and game studios with engine talent where focused on the more standard platforms... and more and more studios where starting to not have in house engine talent at all. Making optimizing for PS3 damn near impossible.

Sure the xbox of the day had a powerpc core... but it was a standard triple core power core that did general computation much like x86. Compiling for that machine was not that hard it was still general compute hardware and MS libraries basically made it a easy recompile.

The upside for Sony was they developed the core of their shader APIs back then... and it forced Sony to add a lot of quality of life things for programmers to their APIs. A lot of game designers prefer sonys APIs over all others today. No doubt the PS3 software development has helped Sony make the PS4 a system developers have been able to optimize very well.
Click to expand...
The API on the PS3 was just a combination of OpenGL ES and Cg. There is no indication at all that GNM and GNMX is an extension of PSGL. The shader language Sony created for the PS4 is completely custom.
 
  • Like
Reactions: ChadD
like this
Armenius said:
The API on the PS3 was just a combination of OpenGL ES and Cg. There is no indication at all that GNM and GNMX is an extension of PSGL. The shader language Sony created for the PS4 is completely custom.
Click to expand...

Of course its not the same API. I mean to say sony took some of the ideas they learned with the PS3 API and extended them. No GNM is not PSGL 2.0 or anything... but Synchronization and performance extensions that where introduced with PSGL mostly due to the PITA to code cell ended up as part of GNM and GNMX. I believe it also spurred them down the dual API road. GNMX is high level and wraps to GNM... meaning most developers end up getting most of the low level API performance without having to really do any heavy lifting. One of the issues with PS3 and PSGL was in order to really optimize you had to get your hands dirty. For the most part Sony has made it a lot easier to do lower level optimization for the PS4.

Anyway yes your right not an extension... but it pointed out what they needed to do to make developers happy. Even before the PS3 sony hardware was always a PITA to deal with. Sony cleaned it up a lot after 3.
 
So what you’re effectively saying is that if the software was written in an optimal way it would be fast. Guess what, you can optimise for x64 too, and intel/amd too.

Speed of computing is directly linked to how optimised software is, just so happens that it is much easier to optimise for x64 than it was the cell.
 
Red Falcon said:
Also, FLOPS isn't everything.
Integer performance is still extremely important, and the Cell is abysmal at integer performance; this is understandable as it wasn't what the Cell was designed for, but that article was garbage...
Click to expand...

Alsoalso FLOPS doing what? Frequently you see specialized chips like the Cell, GPUs, DSPs, etc tout their big FLOPS numbers. However they aren't giving you a real performance picture. They are telling you how fast it can run its best operation on its most favourable data set. Now that's all well and good, because when you build a system with a DSP you are usually doing so precisely because the DSP is good at what you need done. However you want to be careful when comparing that to a CPU. What makes CPUs so useful is they tend to be good at everything. Whatever you task needs, they have hardware to do that well. That does mean that they aren't as fast as a purpose built ASIC, but it does mean they can competently do a wide range of tasks. Thus while a DSP may be faster for particular things, a CPU is usually more powerful overall.
 
Keljian said:
So what you’re effectively saying is that if the software was written in an optimal way it would be fast. Guess what, you can optimise for x64 too, and intel/amd too.

Speed of computing is directly linked to how optimised software is, just so happens that it is much easier to optimise for x64 than it was the cell.
Click to expand...

That is the point of general compute stuff like X86 you don't need to heavily optimize it. It is designed to perform fixed functions.

With the line of argument being used for the cell the same could be argued for GPUs... they are many times faster then CPUs for SPECIFIC things. There is no way you can make a CPU crunch the AI type stuff GPUs are good at. All the software optimization in the world can't make a x86 CPU compete with a GPU for doing simple float point calculations used in real time rendering ect.

Its not that the Cell was a better general compute CPU then x86. Its that if you needed it to crunch low precision math(which games do use a ton of hence the GPU thing) the Cell is superior to x86 design.

The issue is Cell was unlike anything that came before or since... so software wise you have to actually understand the hardware very well to optimize it. Which is asking a lot of a game studio. The Cell really shone in super computer duties where computer scientists optimized the cell to calculate what it was best at. Roadrunner IBMs cell based super computer used 12 or 13 thousand 4 ghz cell chips with an extended resister range and standard DDR 2 mem controllers.

Cell would have made a terrible PC chip... its not a general compute power house. It has one striped down Power PC core.
 
"Far stronger" is a bold choice of words, with out clicking the link I could see it holding up very well against Intel chips at a very specific work load but in general it would get trounced... This seems far far too clickbaity.
 
So what you’re now saying is that it is a low precision powerhouse, sort of like a co-processor for graphics, sort of like what intel have on their processors at the moment... sort of like maths co-processors were in the day...

Where do you draw the line? Technically the GPU sits on (next to) the CPU.
 
Keljian said:
So what you’re now saying is that it is a low precision powerhouse, sort of like a co-processor for graphics, sort of like what intel have on their processors at the moment... sort of like maths co-processors were in the day...

Where do you draw the line? Technically the GPU sits on (next to) the CPU.
Click to expand...

The GPU was really where Sony was looking at the Cell early in the development cycle. You can find talk about Sony doing demos where they show graphics running slow until the Cell is engaged then they are fast. It seems that sometime during development they finally realized that it wasn't going to cut it as a GPU and then went to nVidia for a GPU. This is likely why the PS3 has split main RAM and VRAM. You notice that all the PS2, PS4, PS4 Pro, Xbox, Xbox 360, Xbox 1, and Xbox 1X all shared system/video RAM because it lets you make more flexible use of the limited RAM you have. The PS3 didn't, having only 256MB main RAM and 256MB VRAM, since nVidia didn't have time to work in a redesign of their hardware to do shared RAM.

Basically the Cell was a neat idea but didn't pan out in practice. It wasn't competitive for general tasks against other CPUs, and it wasn't competitive for vector/graphics tasks against GPUs.
 
Ya Sony wanted to use it in a game machine... it didn't need a ton of general compute power. Hence the one Power core. Its almost like people would expect a game console to be doing graphics work or something. :)

Cell itself was also not designed for Sonys Playstation alone. The idea was to use it in tons of things... from routers to super computers. It was used in streaming boxes PCIe extension units high end TVs. In the end the only 2 things that shipped in volume where Playstations... and a couple IBM super computers (you can call it volume when each used 10s of thousands of chips I guess). The cells issue is it was a completely different ISA compared to basically anything else around. If IBM Sony and Toshiba managed to get it into more products who knows what could have happened. Perhaps we would be on Cell generation 5 or 6 now. However it was basically a dead end... cause yes GPUs made a lot of jumps from when the Cell idea hit the drawing board to when PS3 actually launched. The cell design pioneered a lot of new stuff it was basically the first real consumer device SOC. Ya calling it = to a modern x86 is a massive stretch... but = to a CPU in doing things GPUs are better at, yep it still is = to x86 in such things. But ya no doubt even low end GPUs are better then high end CPUs in the same regard.
 
sfsuphysics said:
so developers had to program for A SINGLE HARDWARE CONFIGURATION, and they didn't want to do it... sounds like lazy programmers.
Click to expand...
unless it was a console exclusive with no plans for forward or backward compatibility there is no financial incentive to code and completely optimise for a very specific hardware set. That right there is the main reason that the current Sony and MS platforms are so similar.
 
Stoly said:
Certainly, even the PS1 and N64 took some time for developers to really exploit the consoles capabilities.
Click to expand...

Also I don't think it is fair to just call developers "lazy" if they can't take advantage of something that is complex to take advantage of. When you make an extremely complex system, it is going to be harder for programmers to write code that works well with it. Goes double if not only is it complex, but it is new and unique. I mean look at multiple cores: That has taken quite some time before engines have begun really doing a good job exploiting them properly. It is much easier just to code for a single core. Not only has it taken time, but also better availability of tools. The development environments and libraries these days are far better at helping out with properly using many cores.

So is it any surprise that when a new CPU, that nobody had ever used before, with a new way of doing shit that is really complex comes out, developers have some issues with it? Especially when you learn that the dev tools Sony had for it were rather sparse.
 
It’s all about money:

If it costs me 1000hrs (underestimate) to develop game x, but on console Y it takes 2000 hrs due to hardware weirdness, then it costs twice as much to do it therefore twice as much risk.

If I can get 80-90% of the game for 1250hrs and not use all of the potential, then I can sell that. Sure people aren’t going to be thrilled, but I need less investment to see revenue.

Hardware/software legacy helps here too, cause if I use unreal 4 for dev and my coders have used it before, then it requires less time to develop, same goes for x64 vs cell vs mips vs arm.

If I spend $100, expect $150 return, then I am prepared to accept more risk than if I expect $125 return. But if I spend $200, I expect >$200 return. Am I going to get it if the install base is small(er)?

It’s not laziness, it’s reasonable corporate decisions vs risk.

Similarly:
  • I develop for ps3, it costs me twice as much (2000 buckazoids), therefore I need twice as much (2000 buckazoids vs 1000 buckazoids) to break even, then more to show a profit and pay my coders
  • If I develop for the xbox, and only need 1000 buckazoids to build the same game, or very close to, I only need 1000 buckazoids to break even thus if I make 2000 buckazoids I'm 1000 in the green.
  • If I develop for the xbox and I can pay an extra 50% to get 80% of the game released on the ps3, then it's a no brainer, I'll do it so I have more market to sell to, even if it means I don't use the fancy features of the ps3.
  • If it costs 50% extra to go to PC, I'll probably do that first before I go to PS3, because of the install base

The above also illustrates why consoles resemble each other and resemble PCs these days


What helped the PS3 a lot was the Bluray compatibility, cause then you had another reason to get one aside from games, therefore helping with the install base.

In the N64's favour - Nintendo themselves made some of the games that landed on that platform to get things started, this was probably rolled into the R&D cost of the system as a whole which is why, when it took others to get up to speed with it, it mattered less as there were already a few games available at launch.
 
Last edited:
So...As a developer if I can't code to Microsoft's API, it's just not viable. Sounds exactly like laziness to me. It wasn't that the PS3 ran a custom API, it was that the Xbox ran DX.
 
Mazzspeed said:
So...As a developer if I can't code to Microsoft's API, it's just not viable. Sounds exactly like laziness to me. It wasn't that the PS3 ran a custom API, it was that the Xbox ran DX.
Click to expand...

Well done, you completely missed the point of my post. It’s not about the ability to code. It’s a commercial decision.
 
How do I put it another way?

I’m BigGame Inc and just made 100 million on my last round of games

Install base of consoles (by percentages) is :
60% Xbox
38% PS3
2% other

PS3 requires extra tooling (different tools), and has a different architecture to what I am used to. Even with HAL, I’m going to have to work harder (call it 50%?) to produce a game that exploits the system, cause I am going to have to train my coders, and get hardware etc etc.

So with my 100 mil I go ok let’s go long on this, let’s invest 60% in PS3, 40% on Xbox.

This means I get approximately the same game (with some shiny features on each due to differences in hardware.

I need to sell 50% more games to Break even, so where I need to sell into (call it) 20% of Xbox owners to start making profit, I need to sell into 30-40% of PS3 owners to make profit.

That means that I need to work harder to promote my product, and look at monetisation strategies (loot crates?) in order to make back my investment in PS3, where I can just pretty much coast into Xbox, and still turn a profit.

And guess what? I can use 80% of that Xbox dev on pc, so I can make more money for that investment...

If I was investing my own money, I know which way I would go.
 
ChadD said:
Accept it is for the most part... the PS3 Cell is not however a general purpose CPU, so its not a fair fight.

You could say basically every GPU for the last 10 years is also faster then any current CPU if your taking about doing FP8 vector math calculations. The PS3 was a hybrid... it a had a Power based PPE, it was very close in design to the of the day 64 bit power chips, although it wasn't. (IBM used a version of that core for the xbox) On the PS3 Cell the PPE controlled 6 SPE (synergistic processing units) each SPE had 6 cores with its own embedded L1 cache system... there where some serious issues programming for them as they couldn't directly access system memory among other PITA issues. However in a PS3 cell operating at 3.2 GHZ each unit had a theoretical output of 25.6 Gflops of power. Of course this isn't general compute math... and the Cell also lacked things like branch prediction. The developer in this article is actually spot on the PS3 IF (and that is a big if) is programmed by someone that knows how to setup proper virtual memory maps for all of the SPE units... and they have optimized the hell out of their compiled code to make the lack of branch prediction a non issue you could in theory run software on the PS3 many times faster then MOST modern CPUs, and no doubt many times faster then anything on the market during the PS3 days.

The PS3s combined cell hardware operated at a theoretical max of 230 gflops. For reference not that long ago Intel advertised the Core i9-7980XE as their first teraflop CPU. And the 10980 is basically the same chip. So by that logic if a 18 core Intel chip is only theoretically capable of 4 times the calculation power of a PS3, well 4 and even 8 core versions are at best =. (again we are not talking general compute on a PS3... but it was a very innovative design it was just always better suited to super computer work then game consoles with all the compiler level tricks required to put even a fraction of that power to work in a game) For very purpose built bits of math though Cell arch was unmatched... which is why roadrunner was the fastest super computer in the world till the end of 2009, and by comparison that machine didn't even run that many CPUs. Modern super computers are using 3-4x as many chips in their arrays these days. The only reason IBM didn't continue developing newer Cell based chips was the rise of the GPUs. They basically do the same job.
Click to expand...


I thought the lack of direct memory access was an intentional "feature" added by Sony for the PS3 Linux in order to keep people from playing pirated/homebrew games. If you were in the native console OS, then the software running through that DID have DMA.
Stretching back a lot of years to remember that.
 
Keljian said:
Well done, you completely missed the point of my post. It’s not about the ability to code. It’s a commercial decision.
Click to expand...

He said "viable", referencing a financial, or commercial, decision. He simply didn't need 2.5 books to form his response. That apparently upset you.
 
I may be misremembering, but didn't most devs just use the SPEs for audio processing?

Even Gran Turismo, Sony's flagship exclusive that spends 5+ years in development for each entry, didn't do much with the cell.
 
viivo said:
I may be misremembering, but didn't most devs just use the SPEs for audio processing?

Even Gran Turismo, Sony's flagship exclusive that spends 5+ years in development for each entry, didn't do much with the cell.
Click to expand...
I wouldn’t rank Gran Turismo up that high as the flagship exclusive, despite the development length.

I’d look more at what Naughty Dog managed to do with Uncharted, or Guerrilla with Killzone. I’m not well read on it but I do believe that Naughty Dog put out a fairly informative article on using Cell in the making of Uncharted 2. They go in depth on the differences of how they used Cell between the games too if I recall rightly.
 
You must log in or register to reply here.
Back
Top