Cell: Future of Gaming?

[91TTZ]

lol, I found this article while searching CNN:

PS2 Hype

Read it and take a look at the claims. It's hilarious how they are pulling the exact same stunt they pulled last time. The same hype, the same completely bogus claims.

Remember that this article was printed before the 1 ghz CPU's from Intel came out.

Here are a few snippets:

"the story of the day was Sony's upcoming PlayStation 2, and the Emotion Engine processor that will run it. Developed by Sony and Toshiba, experts predict the high-tech processor will offer unprecedented gaming power."

"that makes the chip two times faster than a 733-MHz Pentium III and 15 times faster than a 400-MHz Celeron at handling tasks like full-motion video."

"It's nothing short of amazing," says Keith Diefendorff

"Diefendorff called upcoming consoles such as the PlayStation 2 a Trojan horse. PC makers don't fear them right now, but they should because he says companies such as Sony "have something in mind other than just games"



So even back then, they claimed to be 2 times faster than the fastest PC's, and 15x faster than an average home PC. Now they are doing the same, except their claims are getting bolder.

 

[Eug]

Sony has a hyperactive PR dept. Get over it. The PR is a lot less interesting than the chip itself.

Anyways, here is an interesting thread about Cell.

It's more technical than I can understand, but it has more specifics than most sites I've seen so far. Linus Torvalds is even in there, although he seems to understand less about the chip than the others in the thread.

Excerpts:

--

1. 90nm SOI process.

2. Logic depth is functionally equivalent to about 20 FO4 (est), but circuit speed equivalence is 11 FO4 per stage. The short pipestage circuit depth is reached with "circuit efficiencies" and Dynamic logic !?!

3. With per stage delay of 11 FO4, the schmoo plots show that the SPE's can crank from 3.2 GHz @ 0.9V Vdd to 5.2 GHz @ 1.3 V Vdd. The entire chip has similar frequency/voltage range, but to get to 5.2 Ghz @ 1.3V, each SPE will eat 11~12W. Add in the rest and the chip will get really hot. 4 GHz @ 1.1V = 4W per SPE seems to be the nominal range.

4. Die size per SPE is 2.5 x 5.81 mm^2. The entire chip with 8 SPE's seems to be about 17.2 x 12 mm^2. That seems to be an awfully large chip for IBM. The CPU to be used in PS3/Xbox2 will probably be the 65nm version or it'll have to have fewer SPE's.

6. As previously announced, the off chip I/O interface is Rambus Redwood and the memory interface is XDR. Similar clocking/deskewing schemes. Looks to be about ~50 GB/s BW to memory, and 50~100 GB/s to I/O.

--

I just spent about 20 minutes with the gentleman who designed the dynamic logic gates for CELL, and he confirmed my suspecion that the PPE (powerPC processing Element) is based on a "new" 64b PPC core, and the lineage of the processor came out of the research papers presented in ISSCC in 2000. He also gave me his analysis/opinion of the scalability of the dynamic logic down to 45 and 65nm. (short summary: It's just more work.... If dynamic stops working, we'll go static, but keep dynamic for the speed/power/area as long as the process can do it.)

 

[91TTZ]

Originally posted by: Eug

It's more technical than I can understand, but it has more specifics than most sites I've seen so far. [/i]

It is more technical than I can understand also, but I do understand one thing that they seem to be overlooking- you can't tell how the chip will run until you actually build it.

By specs alone you can't always tell how it will perform. Remember when Intel designed the P4? They were saying how its design did slightly less work per clock cycle but the design allowed the speed to scale up much, much higher than anything else. They were right about the less work per clock cycle part, but they've seemed to have fallen on their face in regards to ramping up the speed. From what I've heard, they ran into some unexpected problems getting the chip to scale to higher clock speeds. Design-wise it shouldn't have this problem, but things work differently in reality than they do on paper.

 

[Eug]

The P4 ramped up in clock speed far higher than any other mainstream processor. I think they were fairly successful... until they hit the 90 nm wall... No 4.2 GHz P4 or whatever. Still, 3.8 GHz on a P4 is nothing to sneeze at.

I think based on the specs Cell will be speedy. It will NOT be anything like the hype suggests in real world usage on average, but like I said, I find the chip itself more interesting than the hype.

 

[VirtualLarry]

Originally posted by: 91TTZ

Originally posted by: Eug
It's more technical than I can understand, but it has more specifics than most sites I've seen so far. [/i]

It is more technical than I can understand also, but I do understand one thing that they seem to be overlooking- you can't tell how the chip will run until you actually build it.

Yes, and the fact that they have schmoo plots indicates that they actually have test silicon to characterize, as I understand it. It seems very real to me. I think that they are stretching the notion of "cores" in a CPU though, it seems more like a single CPU "core" (the PPE), capable of SMT, and having 8 more vector units (SPEs) slaved to it. My biggest question at this point - are the SPEs capable of control-flow, or predicated execution? Or does that responsibility lay with the PPE exclusively, handing out work to the SPE pipelines to execute, and retreving output work and sending it back to the rest of the system from the Cell? (I mean, if you used Sony's definition of "cores", your ordinary Athlon XP chip would be considered multi-core then. )

Originally posted by: 91TTZ
By specs alone you can't always tell how it will perform. Remember when Intel designed the P4? They were saying how its design did slightly less work per clock cycle but the design allowed the speed to scale up much, much higher than anything else. They were right about the less work per clock cycle part, but they've seemed to have fallen on their face in regards to ramping up the speed. From what I've heard, they ran into some unexpected problems getting the chip to scale to higher clock speeds. Design-wise it shouldn't have this problem, but things work differently in reality than they do on paper.

You're right, but they didn't have schmoo plots out to 10Ghz with the early P4 chips either.

As far as mfg goes, these things really look pretty cutting-edge to me, looks like they've pulled out all the stops. I wonder what the yield rates are though. Power dissapation seems reasonable, although a bit on the higher-end of things if these chips are supposed to be in consoles without an off switch, and potentially embedded into other consumer devices. I'm sure that they probably can power-down individual SPEs and most of the PPE when there's no work to be done though.

 

[zakee00]

Originally posted by: Eug
The P4 ramped up in clock speed far higher than any other mainstream processor. I think they were fairly successful... until they hit the 90 nm wall... No 4.2 GHz P4 or whatever. Still, 3.8 GHz on a P4 is nothing to sneeze at.

I think based on the specs Cell will be speedy. It will NOT be anything like the hype suggests in real world usage on average, but like I said, I find the chip itself more interesting than the hype.

yeah if it dosn't throttle down
the P4 is a pretty bad chip IMO i mean its soooo ineficient. uses like twice as much power as an AMD 64, alot more heat, similar performance (alot worse in some things [games], a bit better in some others [encoding])

 

[xbdestroya]

Originally posted by: zakee00

Originally posted by: Eug
The P4 ramped up in clock speed far higher than any other mainstream processor. I think they were fairly successful... until they hit the 90 nm wall... No 4.2 GHz P4 or whatever. Still, 3.8 GHz on a P4 is nothing to sneeze at.

I think based on the specs Cell will be speedy. It will NOT be anything like the hype suggests in real world usage on average, but like I said, I find the chip itself more interesting than the hype.

yeah if it dosn't throttle down
the P4 is a pretty bad chip IMO i mean its soooo ineficient. uses like twice as much power as an AMD 64, alot more heat, similar performance (alot worse in some things [games], a bit better in some others [encoding])

Well it's not like Cell's going to be built on Intel's process - it's going to be built on the same process AMD's E-stepping chips are going to come out on. That's right, IBM's process.

 

[Terumo]

Originally posted by: MisterChief
That's all fine and good, but I am not a Sony fan. Therefore, I look forward to Microsoft's Xbox Next (Xenon) along with it;s dual/triple IBM processors and ATI chipset. The cell processor may be amazing, but the only way the cell can be utilized to its full extent is if game designers can themselves utilize its unique architecture for gaming purposes.

That's on the near horizon. Gamers were royally upset of what Ion Storm did to the Deus Ex sequel, making it for XBox, then porting it to PC. PC gamers were stuck with a bad game interface and more, and Deus Ex 2 sunk to oblivion.

The PC game market will eventually be taken over by the consoles, which is a shame because it'll reduce the need to built a fancy game rig, when it could be bought for $400 at Wal-Mart.

 

[zakee00]

Originally posted by: xbdestroya

Originally posted by: zakee00

Originally posted by: Eug
The P4 ramped up in clock speed far higher than any other mainstream processor. I think they were fairly successful... until they hit the 90 nm wall... No 4.2 GHz P4 or whatever. Still, 3.8 GHz on a P4 is nothing to sneeze at.

I think based on the specs Cell will be speedy. It will NOT be anything like the hype suggests in real world usage on average, but like I said, I find the chip itself more interesting than the hype.

yeah if it dosn't throttle down
the P4 is a pretty bad chip IMO i mean its soooo ineficient. uses like twice as much power as an AMD 64, alot more heat, similar performance (alot worse in some things [games], a bit better in some others [encoding])

Well it's not like Cell's going to be built on Intel's process - it's going to be built on the same process AMD's E-stepping chips are going to come out on. That's right, IBM's process.

yeah im not sure why i even posted that, just a rant i guess

 

[VirtualLarry]

Originally posted by: xbdestroya
Well it's not like Cell's going to be built on Intel's process - it's going to be built on the same process AMD's E-stepping chips are going to come out on. That's right, IBM's process.

Wow, Intel not being on the cutting-edge of process-technology. Now THAT *is* revolutionary.

 

[RussianSensation]

1.3v @ 85 Celcius operation with heat sink

3 Pages of posts and not 1 person complaining how hot it is??? Come on where are all the Preschott haters?

Also last time I checked 100000000 MIPS and Gigaflops processing power was great, but playing on a 640x480 TV isn't going to make PC users run (and not everyone has HDTVs) for cover. Consoles always make a splash and then 1 year later PC technology catches up and in 3 years PS2 looks like NES.

Either way, if you are truly a gamer, you must have consoles for sports and driving and fighting and platform/RPG games and PC for simulation puzzle strategy and FPS games. There is just no way around it. Can't everyone just be happy?

 

[xbdestroya]

Well that was supposedly a bounds stress test - normal operation is 30W dissipation, or so I hear.

There is so much info on Cell floating around the web right now that I have no idea where I got that from at the moment, all I can say was it was today or yesterday.

So yes, it might be BS, but no, I didn;t make it up.

EDIT: You know, until I find it somewhere, just forget I said it - I myself doubt the stat. But I'm pretty sure they're not going to put 85 C inside the PS3.

 

[Eug]

Yeah, the 1.3 V is supposed to be at 4.6-5.2 GHz in the lab. At 4 GHz it can run at 1.1 V.

 

[zakee00]

Originally posted by: Eug
Yeah, the 1.3 V is supposed to be at 4.6-5.2 GHz in the lab. At 4 GHz it can run at 1.1 V.

i'd LOVE to see that

 

[ribbon13]

Hmmm. A modern GPU has way more Gflops than a modern CPU, but you don't see program code running on it do ya?

Maybe cell based helper processors on a video card for real PCs. =D

 

[VirtualLarry]

Originally posted by: ribbon13
Hmmm. A modern GPU has way more Gflops than a modern CPU, but you don't see program code running on it do ya?

Maybe cell based helper processors on a video card for real PCs. =D

Both of those are actually good ideas.

There is a program that does professional audio-effects processing on high-end NV GPUs now, I forget what it's called. But GPUs can be used for bulk numerical processing. The problem right now is a lack of generalized flow-control, but that's part of how they can run so quickly and efficiently too.

 

[Acanthus]

Originally posted by: zakee00

Originally posted by: Eug
The P4 ramped up in clock speed far higher than any other mainstream processor. I think they were fairly successful... until they hit the 90 nm wall... No 4.2 GHz P4 or whatever. Still, 3.8 GHz on a P4 is nothing to sneeze at.

I think based on the specs Cell will be speedy. It will NOT be anything like the hype suggests in real world usage on average, but like I said, I find the chip itself more interesting than the hype.

yeah if it dosn't throttle down
the P4 is a pretty bad chip IMO i mean its soooo ineficient. uses like twice as much power as an AMD 64, alot more heat, similar performance (alot worse in some things [games], a bit better in some others [encoding])

Flamebait. The pentium 4 Northwood-B and Northwood-C were insanely successful up until the prescott, and prescott sales still outpace Athlon sales even in their current dismal state.

 

[Terumo]

Originally posted by: Acanthus
Flamebait. The pentium 4 Northwood-B and Northwood-C were insanely successful up until the prescott, and prescott sales still outpace Athlon sales even in their current dismal state.

It's due to practicality. Most folks don't want to spend their weekends tweaking their system, like most folks are shade tree mechanics. So most folks will use the plug-and-play processor. Businesses are especially like that to reduce IT costs alone.

 

[ribbon13]

Perhaps you mean hardware tech costs. An IT probably knows how to install a cpu, but main job function revolves around network hardware, not general computer hardware. picking nits bored at 6 in the morning

I'd say Intel sales have a lot to do with Dell and the business sectors reliance on dell.
Boeing for example. Everytime I go to thier surplus store to find a cheap pc... guess what brand it is?

 

[Terumo]

Originally posted by: ribbon13
Perhaps you mean hardware tech costs. An IT probably knows how to install a cpu, but main job function revolves around network hardware, not general computer hardware. picking nits bored at 6 in the morning

IT departments tend to do it all. Not every company is like IBM, where specialization exists. Most businesses hire a couple techs and their job is to maintain everything related too computers. The last thing they would be wanting to do is to OC a PC, while they're trying to keep viruses and other exploits off the network.

Crap, my tech relative is spending hours a day just cleaning networks just before doing any maintence on them. :/

I'd say Intel sales have a lot to do with Dell and the business sectors reliance on dell.
Boeing for example. Everytime I go to thier surplus store to find a cheap pc... guess what brand it is?

Why the wise don't buy Dells. Large companies have money to blow, medium sized companies and mom and pops have a bottom line to meet.

 

[Vee]

I've just taken a second peek at the Cell processor.
The way I see it, the cell can be really fast at straightforward transforming operations, without branching. So it's really a sort of generalized software controlled processor of the signalPU/GPU kind of variety. It can do general computing too, but performance with branching code is poor.
So the idea seem to be to increase the same kind of processing performance, that we use dedicated graphics cards, audio cards and vector extensions - to the instruction set - for.

4GHz 8-cell Cell can theoretically *peak* (not real performance) at ~24GFlops DP. (I think that's achievable with further vector extensions, instruction parallelism and multicore on x86-64 too, if the need was really there.) But that's for a branchless task (compare GPU). I think branches would reduce the Cell to less than a single Gflop, maybe even much less.

- BUT: For 'precisionless' (drop results, no rounding) SP, the Cell could achieve ~200GFlops. Obviously quite useful (understatement) for some special purposes.

The Cell processors impact on computing is going to be really interesting to watch. Sony is obviously going to have massive use for it in everything media appliance. But don't count on it to replace x86-64 any time soon.

I'm tempted to see it more as a sort of programmable GPU, with cpu functionality, than a real CPU. Which asks the question about how tasks are going to be divided between PS3's nVidia chip and the Cell. It would surprise me greatly, if Sony do not intend to let the Cell do lot's of transformation stuff that is normally done inside the GPU. Interesting here is the flexibility the Cell offers, since its processing is entirely software controlled.

Also interesting to note is that nVidia have been making some noises about the idea to accomplish a similar software controlled 'generalized' GPU architecture. The kind of rendering such a videocard could do, would solely depend upon the software. It could even conceivably do radiosity and raytracing, for an architectural app, or "real 3D" for a slowmoving exploration game like Myst.
The Cell processor could bring some fascinating capabilities to the PS3. I think I'll have to get one, for once. (don't own any console).

I think what we're basically seeing here, is the option of having dedicated hardware accelerated processing (audio, video, anything similar), go back to software.
Sony will have a general hardware chip to massproduce and run EVERYTHING on!
Everything else will be software.

So yes, the Cell should be exciting to watch.
Future of gaming? Maybe. At least the concept as such, but I don't think ATI and nVidia are just going to roll over.

 

[VirtualLarry]

Originally posted by: Vee
I've just taken a sekond peek at the Cell processor.
The way I see it, the cell can be really fast at straightforward transforming operations, without branching. So it's really a sort of generalized software controlled processor of the signalPU/GPU kind of variety. It can do general computing too, but performance with branching code is poor. So the idea seem to be to increase the same kind of processing performance, that we use dedicated graphics cards, audio cards and vector extensions - to the instruction set - for.

Interesting. Can you say whether or not the Cell's SPEs (not the PPE) have their own flow-control capability? Or is it controlled entirely by the PPE? It would seem to me, that if the SPEs are designed to be slaved to the PPE "master", then the ability for the SPEs to execute flow-control instructions on their own, would be undesireable, because it would make their execution times non-deterministic, which would interere with the PPE's proper scheduling of being able to feed the SPEs with data. But perhaps I still don't fully understand the architecture.

It's kind of interesting though, because if you recall how the original 8086 and 8087 operated - they were in fact independent co-processors, attach to the same bus, and the 8086 handled all integer and flow-control instructions, while the 8087 interpreted and processed the floating-point instructions. But the 8087 lacked any sort of flow-control instructions on its own, and the opcode stream had to have FWAIT opcodes inserted into it periodically to sync up both independent co-processors, in order to operate properly.

Originally posted by: Vee
4GHz 8-cell Cell can theoretically *peak* (not real performance) at ~24GFlops DP. (I think that's achievable with further vector extensions, instruction parallelism and multicore on x86-64 too, if the need was really there.) But that's for a branchless task (compare GPU). I think branches would reduce the Cell to less than a single Gflop, maybe even much less.

In other words, when forced to execute general-purpose (heavy/non-deterministic branch-filled) code, the performance of the entire Cell drops to around the performance of the PPE itself, I would assume.

Originally posted by: Vee
The Cell processors impact on computing is going to be really interesting to watch. Sony is obviously going to have massive use for it in everything media appliance. But don't count on it to replace x86-64 any time soon.

Exactly! It sounds very exciting, and performance-oriented... but only for rather specialized tasks.

Originally posted by: Vee
I think what we're basically seeing here, is the option of having dedicated hardware accelerated processing (audio, video, anything similar), go back to software. Sony will have a general hardware chip to massproduce and run EVERYTHING on! Everything else will be software.

I've suggested several times in the past, to create a "soft GPU", comprised of (at the time) multiple Athlon XP CPUs with a shared memory pool, although an SMP cluster with AMD64 chips, each with their own dedicated high-speed DRAM pool would be a better solution. But the advantage would be that it wouldn't be so obsolete with each generation's changing of the 3D API standards. In fact some cards could be built as entry-level, with some of the CPU/DRAM slots unpopulated, and you could add them later, ala SLI almost.

Maybe I'm a bit biased in that regard, I once worked for a game company, with some Brown Uni. grad guys, that had a software rendering engine running, that was faster and more feature-filled than the (mostly non-existant) 3D PC hardware accelerators of that day. Software is infinately more flexible, and given the right mix of hardware opcodes, can be nearly as performance-efficient as some of the hardware solutions. It's really a cost-performance see-saw, there is no inherent benefit to implimenting a rendering engine in either software or hardware, it all comes down to cost/performance ratios at the time.

 

[Vee]

Originally posted by: VirtualLarry
Can you say whether or not the Cell's SPEs (not the PPE) have their own flow-control capability? Or is it controlled entirely by the PPE? It would seem to me, that if the SPEs are designed to be slaved to the PPE "master", then the ability for the SPEs to execute flow-control instructions on their own, would be undesireable, because it would make their execution times non-deterministic, which would interere with the PPE's proper scheduling of being able to feed the SPEs with data. But perhaps I still don't fully understand the architecture.

(sorry, I've been busy a few days, and I'm off again tomorrow)
Yes, as I understood it, the SPEs have their own flow-control. It's very straightforward though. No fancy logic tricks, beyond load/store forwarding. Just 'in order execution', and no branchprediction features visable. I don't think it can do speculative execution.

In other words, when forced to execute general-purpose (heavy/non-deterministic branch-filled) code, the performance of the entire Cell drops to around the performance of the PPE itself, I would assume.

That was my speculation. However I made that estimate based on the lack of branch handling features. Also the SPEs can run really fast only in their own 256Kb LS space. Another reason any branches should prove to be disastrous for performance.

The SPEs are there to apply an operation onto a set of data. They would suck at any type of analytical computing. (much as P4's implementation of SSE2).

There are some murmurs in 'highly technical' about that the Cell processor should be first seen as a general CPU, since nVidia is supplying some kind of graphics chip for PS3. However, the Cell processor cannot display any graphics, so obviously some "graphics" chip is needed. And the borderlines between what cpu and gpu each contribute to imaging, is in no way fixed. GPUs have gradually taken on more. There is no reason why that process cannot go the other way too. I will be surprised if the Cell will not have something to do with image rendering in PS3.

Edit: Finally found the link to a good article, here you go:

http://www.realworldtech.com/page.cfm?ArticleID=RWT021005084318