Some quotes about PS3 CPU

[nAo]

I'm transcripting these quotes from PS3Insider's Forum cause I believe this stuff is very interesting
All these quotes come from the same guy (his name is cpiasminc) from several posts:

I did go visit another SCEA office yesterday, and one of the guys said that the samples they had (again 1.15 GHz, I assume -- they didn't say) achieved a peak of around 9 GFLOPS to each APU. Now if I assume that it is 1.15 GHz, that means that eiher the APU is 256-bit (and can therefore do 8 32-bit FLOPS per cycle), or it is 128-bit and is a wide issue multi-pipe APU and can issue up 2 SIMD instructions per cycle (which would be better). Unless it supports 16-bit floats as well, which I don't know why they would. Anyway, that would yield 8 FLOPS per cycle and therefore 8 * 1.15 GHz = 9.2 GFLOPS. Sounds about right. Unless this office has double the clock of the mock-ups I saw before

(note: he went to SCEA offices a couple of times..first time he saw a test board clocked at 1.15 Ghz. When he went over the second time he was told about the 9 gigaflop/s per APU figure)

Well, the engineering samples I saw were running pretty darn fast. They were only 1.15 GHz, but the guy did say they were pretty old (consider the fact that these parts happened to ship to California...) At the 1.15 GHz level, the wattage was only around 20 W running Soul Reaver inside the still buggy PS2 emulator. However, the actual power consumption will go up as clock speed rises, and the practical consumption will go up as your code better utilizes the processing facilities.

...

Everything was pretty encouraging except for sound. None of the mock-ups had sound yet. The guy said that they Sony have to evaluate some costs in order to make any decision about the final audio arrangement. I did have to sign an NDA so I'm not allowed to say much, and they were also kind of vague about it. I got the impression that they're not even sure where to put it in the pecking order. Considering that cost was brought up, I'd have to wager that there may not be a third Cell-like chip that uses DSP-based APUs for sound. It may just end up with an off-the-shelf audio controller.

Developers only have cycle-counting emulators. So their tests run at around 0.2 fps or so, but there's a little clock-cycle estimator in the corner. And apparently since the scheduling/distribution logic isn't so well emulated, the clock cycle estimator is said to be correct within 5-10%, which is basically really bad.

I still have doubts about the schedule. They gave me no guesses regarding how on-track they are and they have no info about how far things are in Japan.

....

BTW, there was news a long time ago of CELL prototypes running at 200-800 MHz, and those were of all sizes ranging from 4 APUs to 128 APUs

ciao,
Marco

 

[McFly]

128 x 9.2 = 1177.6GFlops or about 1TFlops ... with only 1.15GHz, so if they can go up with the clock and I guess this is only the BE ... holy shit!

Question is if they can go up that much with the 128 APU version, but it seems like 1TFlop is not impossible if he speeks the thruth.

Fredi

 

[Megadrive1988]

128 APUs would fit nicely with reports from a long time ago (2+ years i think) that PS3's main processor would have upto 16 CPUs in it. (4-16 processors, remember?) That would be 16 PowerPC cores, or whatever cores they decide on. so then we would assume that each of the 16 CPU cores would be in charge of 8 APUs.

sounds crazy but who knows....

 

[jvd]

128 APUs would fit nicely with reports from a long time ago (2+ years i think) that PS3's main processor would have upto 16 CPUs in it. (4-16 processors, remember?) That would be 16 PowerPC cores, or whatever cores they decide on. so then we would assume that each of the 16 CPU cores would be in charge of 8 APUs.

sounds crazy but who knows....

and u really think the die size would be reasonable on 65nm process ?

 

[Megadrive1988]

like i said, it sounds crazy.

I think 8 PPC CPUs with 64 APUs might be possible though.
giving ~512 GFLOPs at around ~1 GHz.

 

[Guden Oden]

I don't expect 1GHz CPU in the 06-07 (euro launch) timeframe; too slow. XB runs at 733MHz now, only 266MHz increase in all these years will look very bad to Joe Public who don't have a clue about how performance really is measured.

 

[AlNom]

I don't expect 1GHz CPU in the 06-07 (euro launch) timeframe; too slow. XB runs at 733MHz now, only 266MHz increase in all these years will look very bad to Joe Public who don't have a clue about how performance really is measured.

They could always go for the 100 gigabillion operations per second

 

[kenneth9265_3]

so for us non techicial people...we are talking about a really powerfull gaming console....Right? :?

 

[jvd]

I don't expect 1GHz CPU in the 06-07 (euro launch) timeframe; too slow. XB runs at 733MHz now, only 266MHz increase in all these years will look very bad to Joe Public who don't have a clue about how performance really is measured.

Mhz isn't going to matter .

Once sony starts its hype fans will be throwing around numbers they don't even understand. Justl ike the launch of the ps2 .

 

[Paul]

I don't expect 1GHz CPU in the 06-07 (euro launch) timeframe; too slow. XB runs at 733MHz now, only 266MHz increase in all these years will look very bad to Joe Public who don't have a clue about how performance really is measured.

Bingo.

 

[MfA]

Why are they showing this stuff to people who dont grok fused multiply adds?

 

[passerby]

Now nAo, as a respected member of these boards, I hope you are making reliable references! Anyway is there any reason why you think this poster is believable?

 

[Panajev2001a]

Why are they showing this stuff to people who dont grok fused multiply adds?

For the same reason they allowed the GS to be designed withut the supervision of some group of top paid 3D graphics programming gurus.

 

[Tysan]

Ok, so Cell can have between 8 and 128 APUs. But how many of them will Ps3 Cell have?

 

[sunscar]

There's more to Cpia's posts than that, by the way. Keeping in mind that these are barely prototypical, each APU was reported to be capable of around 9 gigaflops. With a likely embodiment containing 32 APUs in the CPU and 16 in the GPU, we get 441.6 Gflops peak for the two units combined. That is if the two units are both running at 1.15GHz. If they're running faster, then we get more, if they're running slower, then less.

Admittedly, 128 APUs on a 65nm die doesn't sound all that realistic, but 32 or 16 sound considerably more likely.



BTW.......



To: Passerby

Cpaisminc would be one of those accademian graphics gurus Panajev was referring to, so I think you can consider him to be a viable source.

Anyway, getting late here, so I've got to fly.

Good evening,

Iridius Dio

 

[Vince]

Why are they showing this stuff to people who dont grok fused multiply adds?

Or understand what the term tape-out signifies for that matter. Somehow, I think the probability of what you asked is about the same of *other claims* I've heard such as Sony's CTO giving out the specs during a conference to a random person in conversing... eg. near zero.

 

[V3]

1.15 GHz, 2.85 GHz to go.

I don't get his explanation though, I think he overcomplicates thing, but the APU from the patent would give around 9 GFLOPS with that clock speed.

If the engineering sample is a BE like in the patent, that means its delivering around 288 GFLOPS. Assuming that, the wattage, they need to get it down from 20W when its only running Soul Reaver. If not the chip will be breaking 100W barrier when it is pushing 1TFLOPS+.

 

[cthellis42]

Mhz isn't going to matter .

Once sony starts its hype fans will be throwing around numbers they don't even understand. Justl ike the launch of the ps2 .

You realize that fandom only really sells numbers to itself, right? The mass public who actually buys the machines would be hard-pressed to name an numbers (even RAM rarely gets mentioned), but instead relies on recognizable features (like "it has a DVD drive instead of a CD drive!" or "it has a hard drive!" or "it has built-in THUSANDSUCH!"). They might catch a whiff of excited fans and store owners and magazine readers saying something, but all it translates to is "the XXX is more powerful" or simply "better."

And heck, that's reinforced by release date more than anything else. "It's newer" so therefore "it's more powerful" and "it's better."

The only people babbling about (or even recognizing) any numbers do it on these message boards. For most of the buyers, they're more likely to lean towards the "most recognizable brand" than anything else.

 

[one]

I'm more impressed about PS2 emulation on Cell. Is it emulated entirely by software, or only its GS part is executed on Visualizer? I'd thought PS2 backward compatibility would be done by a shrinked EE I/O processor hardware, and PS1 on emulator software.

 

[nAo]

Why are they showing this stuff to people who dont grok fused multiply adds?

That guy knows FMACs, I had no time to pick up all the 'interesting' quotes.
Moreover I know you guys here are so smart there is no need to quote basic stuff