The PlayStation 3 will feature the much-vaunted Cell processor, which will run at 3.2GHz, giving the whole system 2.18 teraflops of overall performance. It will sport 256MB XDR main RAM at 3.2GHz, and it will have 256MB of GDDR VRAM at 700MHz. It has 6 USB ports, 3 ethernet ports and Bluetooth (no WiFi support for gaming it seems and Bluetooth was chosen for controllers only as it has better latency over WiFi). There are also CF/SD and MemoryStick readers integrated.
Sony PS3 Specifications Are Out
About The Author
Eugenia Loli
Ex-programmer, ex-editor in chief at OSNews.com, now a visual artist/filmmaker.
Follow me on Twitter @EugeniaLoli
i thought i had been following the ps3 thing for quite a while, i never new it was suppose to hav 4 cells… not that i t hink it matters that it doesnt hav them. its suppose to b twice as fast as the xbox 360 and that thing has 3 cpu’S. y would u need 4 cells in 1 box? just hook up yur computer, ps3, tv, and toaster when sony starts making them and u will hav 4 cells interconeected.
if u dont want yur posts deleted dont use any foul languauge, for sumone reason even undirected discriptions with ‘cuss’ words get deleted around here.
I don’t think that the Linux kernel can take advantage of the “cell” technology just yet … so this is pointless.
tru but i bet it will b able to do so eventually. and will we ever b able to build a PC based on the cell CPU? theres still a million unanswerable questions. so back to my point, it will be neet to run linux on it at sum point.
I want to see how this Cell chip handles encoding a DVD, when the software is available. It might not be available right now, but neither is the PS3. Other than that I’d like to use it as a build server, too. But I can wait until its ready.
2 Ghz compile/encode boxes work fine for me now. But I imagine they wouldn’t keep up with an encoder designed to take advantage of the Cell’s capabilities. But who knows, by then we might have desktop Cells.
After seeing both video press conferences of PS3 and XBox 360…I wouldn’t even bother with the MS product.
The PS3 has one word to describe it…Incredible.
The Cell CPU covers the Physics, AI, etc. It even has enough grunt to do software rendering of graphics. It seems to have so much excess power left over, all the demos shown were done in realtime. (To prove, they use the controller of the Playstation to pause and pan/zoom the view of the demos).
And then there’s the Nvidia RSX GPU, this is simply the cherry on top for the PS3. If this GPU is in the next generation video card for PC, I’m not even bothering with GF6xxx series of video cards. Its a quantum leap in lifelike graphics.
And since this is OSNews the PS3 uses is called DNA.
From the presentation, the RSX GPU uses OpenGL and Nvidia Cg APIs. If you’re a PC game developer, your knowledge and experience in the PC industry also applies to the PS3. The transition is relatively smooth, as EPIC developed a demo using the Unreal 3 engine within 2 months of receiving the PS3 Development Kit. (So, it seems to be much easier to develop on).
The Cell can decode 12 High Definition video streams simultaneously, you can have all 12 display and simply select the one you want as the main screen.
The dual display function is cool. eg : You can browse the web and play a game at the same time.
You can play the game on first screen and receive statistical and tactical data on the 2nd screen.
Another application is playing a game and have a video chat with friends while playing!
The iToy from the PS2 is fully compatible with the PS3. Sony demo’ed it with the creator of iToy using it to control plastic cups to scoop up water (in the virtual world) and such.
I am clearly impressed by the PS3. After a long slouch of watching the PC CPU industry literally grind to a halt due to 90nm issues in 2004, this is a very refreshing change.
And besides…I can’t wait for Metal Gear Solid 4!
Forgot to add…The Nvidia RSX CAN access the main memory of the PS3…So it has a total of 512MB of usable RAM. 256MB local memory and 256MB XDR system memory.
@Hiryu:
These specs are pretty outrageous.
However, I’m reminded of how the PS2 was supposed to push 70 million polygons per second with AI and everything else. The PS2 has never come close to that mark.
Of course it has. 3D demos done in LINUX on the PS2 easily hit 30M polygons/sec. Current PS2 games push real close to 70M.
@Myren:
the ps3 was supposed ot have 4 cell processors! 4, not 1. all the various rumors and hearsays have all said 4, for the past x years we’ve been digging.
No, some idiots misinterpretted the old info. The Cell in the PS3 was originally going to have four SPEs. They have since doubled that to eight. It was never going to have four Cells.
MS is screwed with the Xbox 360. The PS 3 is more than twice as powerful and to be released only half a year later with many more AAA launch titles.
Can it run Mac OSX?
Once someone ports Mac-on-Linux to it.
I think I don’t undertand the measure (flops) correctly. If the ps3 can achieve 2 teraflops, you could pile up 35 of them and achieve the performance of IBM’s Blue Genie? Or not?
The form of the PS3 is similar to Nokia Media Terminal from 6 years ago.
http://www.linuxdevices.com/articles/AT4370516520.html
thats overall system performance. the gpu makes up most of it at like 1.8tf or something like that. the difference being that the blue genie has 65(i cant remember the number..) teraflops of GENERAL purpose cpu speed! the gpu isnt general purpose.
The measure of FLOPS is really dependent on the vendor, and also on the target application.
First, there is a difference between the capability of the hardware, and how much is directly accessible. In the Blue Gene, all 65 teraflops is directly accessible to software. In the Cell, easily 70% of the total floating-point operations is tied up in fixed-function circuits in the GPU. Floating point operations are happening, but they aren’t general-purpose.
Second, all FLOPs are not equally precise. The “floating” part isn’t the same, so to speak. The Blue Gene’s processors do fully general double-precision (64-bit) operations with proper IEEE754 semantics for rounding, etc. No unit on the PS3 can do this sort of operation. The VMX unit on the PPE can do single-precision IEEE754 math at a rate of about 12 gigaflops, while the 8 SPEs can do single-precision non-IEEE754-compliant math (it doesn’t support the proper rounding modes) math at 218 gigaflops. The SPEs can also do double-precision non-IEEE754 math, but at a 10x speed hit, which is a still-decent 20+ gigaflops.
Now, the lack of double-precision IEEE754 math isn’t a problem for the Cell’s target market, because most media apps function fine with single-precision math. However, scientific computing apps often need double-precision math, so if your application is like that, the Cell’s 218 gigaflops isn’t very usable.
Third, not everyone measures “operation” the same way. Some people call an FMAC 1 operation, some people consider it 2 operations. Some people consider SIMD instructions to be a single operation, others consider them multiple instructions.
Hope that made things clear as mud
Alright, so… the new “RSX” is more powerful than two 6800s… and two 6800s are about 100Gflops, total…
Can somebody please explain to me:
a) How 100Gflops = 1.8Tflops,
b) who the hell thinks they know how to measure Gflops in consoles, since the current XBox GPU is claimed by the company at 95Gflops, which is almost 2x the raw power of the 6800… which isn’t likely… and I dare say isn’t possible since it’s 5 years old
c) How they can possibly make the claim that this will have more “overall power” than all but 80 of the most powerful computing clusters in the world (see top500.org )
or d) what part of the definition of Gigaflop I am missing, as apparently it no longer means “One billion floating point operations per second.”
They claimed the CPU in the XBox did 2.93 Gflops. I have two problems with this.
The Itanium 733 only does 2.13Gflops.
A stock P3 733 has less than 1Gflop.
Where the hell do the numbers these companies claim come from? Yes, the stock x86 CPUs are weak in floating point operations, but even fixing that the CPU should not even come close to the performance of an itanium, let alone beat a 2Ghz P4 by 2x and an itanium by 1.5x?
Now, is it possible that one PE on a cell is 3x as powerful as a P4? Sure, the old alpha CPUs were 3-4x as powerful as x86 hardware of equal speed… nothing particularly new there…
Is it possible that the GPU is 29x as powerful as the current GPU powerhouse on the market (Radeon X800 XT at 63Gflops)?
I’m gonna take a wild guess and say no, no it’s not.
You can design a custom CPU for a gaming system without the common desktop CPU parts and crank up performance pretty easily, but a GPU is still a GPU, and still has the same job in a game console, cutting corners to boost performance would be a little odd… right?
But I’d really love for somebody to explain to me why I’m wrong?
Thanks helf and Rayiner Hashem for your answers. I didn’t even thought about the floating point presicion or the general purpose registers. I guess I need to do some reading of cpu architecture in the near future. A bit confusing at first, but seems to be a very interesting topic nevertheless
Can somebody please explain to me:
a) How 100Gflops = 1.8Tflops,
First, the “2 Geforce 6800 ultras” is from a journalist, so don’t trust it, at all. The 1.8 teraflops number seems much more realistic. A GeForce 6800 Ultra pushes close to 250 gigaflops, and this thing isn’t coming out for more than a year.
c) How they can possibly make the claim that this will have more “overall power” than all but 80 of the most powerful computing clusters in the world (see top500.org )
Because they don’t measure FLOPs the same way, because the FLOPS aren’t use the same way? Read my post (the one above yours), it clarifies things. Basically, you can’t do molecular dynamics using FLOPs from the GPU, but you sure can do graphics using FLOPs from the GPU. Sony cares about the latter, not the former. Supercomputer manufacturers care about the former, not the latter.
d) what part of the definition of Gigaflop I am missing, as apparently it no longer means “One billion floating point operations per second.”
As I said in the above post, the difference is in the definition of “floating” (how precise, what format, what semantics?), and in the definition of “operation” (is a SIMD add 1 operation or 4, what instructions are available?)
The Itanium 733 only does 2.13Gflops.
A stock P3 733 has less than 1Gflop.
You’re confusing peak theoretical rates with achieved theoretical rates. The 2.13 gigaflops number is an achieved rate. The 2.93 gigaflop number is a theoretical rate (733 * 1 4-component SIMD = 733 * 4 = 2932 mflops). The Itanium gets much closer to its theoretical rate than the P3. Again, as I said, even supercomputer manufacturers cite theoretical rates in their marketing.
Is it possible that the GPU is 29x as powerful as the current GPU powerhouse on the market (Radeon X800 XT at 63Gflops)?
Again, theoretical peaks vs sustained rates. NVIDIA is citing 1.8 teraflops for its new GPU. NVIDIA also cites 250 gigaflops for its Geforce6 chips. Is 7x, a year down the road, believable, with a 300+ million transistor chip running at 550 Mhz? Yes!
Actually, I was talking about the price and pathetic restrictions on the official SDKs
I’m not speaking theoretical rates, I’m speaking average performance.
The 6800 is only about a 53Gflops CPU, it it nowhere near 250Gflops. Even 53 may be giving it a bit of an extra “best case” edge. Nvidia claims the 6800 Ultra at 40Gflops and doesn’t list any “maximums” anywhere…
I’m not sure where you got the “nVidia cites 250Gflops” thing from, however, as nVidia does not generally publish such thing, and many many tech sites have stated the maximums of the cards as “Unknown and unpublished.”
Regardless, lets take a look at your “next year” argument:
Again, theoretical peaks vs sustained rates. NVIDIA is citing 1.8 teraflops for its new GPU. NVIDIA also cites 250 gigaflops for its Geforce6 chips. Is 7x, a year down the road, believable, with a 300+ million transistor chip running at 550 Mhz? Yes!
That is an incredibly stupid argument, and this is why. Five years ago the “Theoretical Maximum” of the XGPU was ~100Gflops, and in the past 5 years that “Maximum” has only increased 2.5 fold (Assuming you know what you are talking about on the 6x GPU, which I’m not too sure on).
Now, you have just suggested that in the next year nVidia will magically make a GPU that has a maximum power 7.6x that of the current power champion in their production line (which nVidia themselves says in not at all normal, see below).
Do you see anything wrong with that assumption? Anyone? Not all at once now…
So, assuming we see the same rate of advancement, I’d guess we can expect something with a maximum of ~320-400Gflops. Or, we can go completely off the wall here and say it will might be 2-3x as powerful, which still puts it around 500-750Gflops?
It is not a sane argument to say that a company can make 760% improvement in one year after 5 years of near stagnation, sorry…
Basically, you can’t do molecular dynamics using FLOPs from the GPU
Actually you can, sorry to tell you. There are many projects around right now working on ways to do just that…
nVidia themselves has various reports that suggest just such a thing is a good idea. Check out http://download.nvidia.com/developer/presentations/2005/I3D/I3D_05_…
(Note nVidia states 6800 was tested at 40Gflops, which is measuring it in the same way as you would measure the 3Ghz P4 in the same presentation)
They seem to be saying exactly the opposite, that GPUs are precise and fast enough to be used for other things…
Also check out http://www.gpgpu.org
Quite a lot of things that would require very heavy/accurate FP operations seem to be there.
Nvidia also states the annual growth rate of a GPU to be 2x, which again puts us at 500Gflops, not 1.8Tflops. Straight from the horses mouth there…
You’re confusing peak theoretical rates with achieved theoretical rates.
I’m not confusing anything, I asked where the number came from given that the real performance of such things was so much lower than they claimed. You’ve given an acceptable reason, which I would have seen earlier had it been there when I started writing my clip… yes, there is a 20 minute difference in posts, but that’s just because I spent a lot of time trying to find things such as the “theoretical maximum” on current GPUs (which is nowhere to be found, please point me to where you got that, I’m curious).
even supercomputer manufacturers cite theoretical rates in their marketing.
Indeed they do, however Intel does not make a habit of it. If you look on their site, Intel will tell you the rating that they themselves measured on their CPUs, not a “theoretical maximum”.
I would also like you to note the difference between what I have listed and what a supercomputer manufacturing company might list.
The things I listed come to ~25-35% of their “Maximum” power, which means that the maximum on them is hogwash, not useful for anything.
A supercomputer can generally be expected to reach 65-95% of their listed maximums. That’s a very large difference there, and as such it makes that suggestion pretty useless as the numbers of the two worlds do not line up.
It’s like trying to tell someone something about the Heisenberg uncertainty principle while they’re talking about tire friction, they are two different worlds and don’t really apply to each other in any way.
I don’t know about that. PS2 managed well, even if Xbox and Gamecube were better machines.
It’s all about the games and image.
What is most interesting about the three ethernet ports is that one is classed as an input port, the other three as output ports. Sounds like it includes a router. Curious.
I think the big question is when are Rockstar planning to release GTA for this machine?
The PS3 is backwards compatible with their product line, but is the Xbox360 backwards compatible, can I run my Xbox 1 games?
I wonder if the blue ray technology is going to catch on, either way you probably want to have a super sized projection television for the PS3 or Xbox 360.
I imagine that the whole setup will cost about $4000.00 USD.
It looks like they are going to be pumping out a lot of games for the PSP. That might justify the price of the PSP for more customers. Sony is strong in the games area.
I’m not speaking theoretical rates, I’m speaking average performance.
And the vendors aren’t speaking about average performance — they are talking about theoretical rates.
The 6800 is only about a 53Gflops CPU, it it nowhere near 250Gflops.
That’s the performance if you consider the maximum throughput of the shaders. It does not include all the FLOPs that are part of fixed-function circuits.
I’m not sure where you got the “nVidia cites 250Gflops” thing from, however, as nVidia does not generally publish such thing
IIRC, the number (200-250 gigaflops, I forget), is from either one of the Geforce6 tech intro documents, or one of the “general purpose programming on GPU” documents. I’ll see if I can dig it up. In any case, ATI *does* claim 200 gigaflops for the X800XT PE (which’d put the X850XT PE somewhere north of that). It can be inferred that these numbers apply fairly well to the 6800, since they are similarly powerful cards.
http://www.cyberscholar.com/ATI/doc/RADEONX800SeriesBrochure129-405…
It is not a sane argument to say that a company can make 760% improvement in one year after 5 years of near stagnation, sorry…
“Near stagnation”? NVIDIA’s GPU Jackpot presentations show gigaflops ratings quadrupling from the NV30 to the NV40 in a year and a half (the I3D presentations show that these numbers 10 gflops and 40 gflops respectively, are measured performance from a specific shader, not theoretical performance for the whole GPU). If you consider this rate in terms of the 2+ year gap between the RSX and the Geforce6, then a 7x increase in *theoretical* maximums doesn’t sound so outrageous.
Actually you can, sorry to tell you. There are many projects around right now working on ways to do just that…
It really depends on what your molecular dynamics code needs. If it needs double-precision IEEE754 math, then no, you can’t do molecular dynamics using the GPU. In fact, for many of the 200 gigaflops in the GPU, you can’t do anything other than the single function (eg: perspective divide), that a particular circuit was designed to do. My point is that the 200 gigaflops number and the supercomputer teraflops numbers aren’t directly comparable. While all those operations might be happening, they aren’t usable in the same way.
Basically, I’m referring to your point about “claiming the PS3 has as much power as all but the top 80 supercomputers”. It’s a lot easier to build an FPU that just does perspective divide than it is to build one that does full double-precision IEEE754 math. If you’re building a supercomputer, you need the latter. If you’re building a console, you can get away with the former. For a console game, those perspective divide ops are useful, needed operations, so its not shady to report them in your total ‘2 teraflop’ figure. However, it’s that difference in application requirements that make the figures between consoles and supercomputers totally incomparable.
(Note nVidia states 6800 was tested at 40Gflops, which is measuring it in the same way as you would measure the 3Ghz P4 in the same presentation)
No! It says it right on the presentation — the P4 number is a theoretical peak, while the GeForce 6800 number was derived from the sustained performance of a shader program that consisted of MUL instructions. Again, note also that they are just referring to the performance of the shader units — not the whole GPU.
Quite a lot of things that would require very heavy/accurate FP operations seem to be there.
Accurate = double-precision floating-point. Lot’s of things are on that list precisely because they *don’t* require “accurate” computation. They can get away with “good enough” computations.
Nvidia also states the annual growth rate of a GPU to be 2x, which again puts us at 500Gflops, not 1.8Tflops.
You forget the GeForce6 is more than a year old now. That puts us at 1 teraflop, which I’d argue is in the ballpark.
I’m not confusing anything, I asked where the number came from given that the real performance of such things was so much lower than they claimed.
The numbers come from the same place all such theoretical numbers come from. You take the number of FPUs in the chip, and multiply by the clockrate. Where do you think BlueGene/L’s 360 teraflop figure comes from? The system consists for 65535 dual-core PowerPC 440 processors running at 700MHz. Each core contains 2 FPUs, each of which can handle a 64-bit FMAC in 1 cycle. Since an FMAC is counted as two operations, the total performance is 2 * 2 * 2 * 65536 * 0.7 = 367001.6 gigaflops = 360 teraflops. This is exactly how IBM got the 218 teraflops figure for the Cell, and likely how NVIDIA gets the 1.8 teraflops figure for the RSX. Note that nobody counts FLOPs the same way, but in the target market of each product, their particular definition of FLOP makes sense.
The things I listed come to ~25-35% of their “Maximum” power, which means that the maximum on them is hogwash, not useful for anything.
Only if you use them for something they were not intended to be used for! If you use the GeForce6 to do CFD calculations, except that you’re only going to get 25% of its theoretical performance, and indeed, the 50 gigaflops figures for things like gpgpu support that. However, if you use the GeForce6 for graphics, then you *are* getting you 200 gigaflops worth (or some 70-80% thereof), if you also factor in all the operations from the perspective divides, triangle setup, etc, that the GPU needs to do in addition to those 50 gigaflops of shader operations.
It’s the same thing with Cell. If you write code that is properly parallel, you’ll get your 200+ gigaflops. It’s no different from a supercomputer, really. A game engine on BlueGene/L isn’t going to give you 360 teraflops (hell, it’ll probably run at the speed of a 700MHz PPC440). In fact, code that gets 360 teraflops out of BlueGene/L should be very similar to code that gets 218 gigaflops out of Cell, because both machines consist of numerous simple, fast, compute nodes with fast local memory connected by a fast interconnect.
This is AMAZING! I’m getting a unit!
Every 5 years the game companies announce new system specs before the finalized product, and gullible people everywhere eat it right up.
“The N64, rivaling the fastest graphics supercomputers for only $250.”
“The 3DO M2, 1 million polygons per second.”
“The PS2 with Emotion Engine(tm), revolutionizing home entertainment – the PC will be obsolete.”
Keep lapping it up, dopeys…
And just like clockwork, every 5 years, a 28 year old living in his parents basement decides he needs to flex his superior intelligence by commenting on how stupid everyone else is. God bless you bearer of desolation, may you grace our paths with your insightful wisdom in another 5 years….
The price of real “230 gigaflops” is around $100K today. Just to give you guys an idea about how much BS the 2000 gigaflops PS3 is.
http://www.newscientist.com/article.ns?id=dn7340
i hate the stupid xbox 360 its lame ps3 smashes on that stupid thing.