How did Sony get 1.8 TFLOPS for the RSX?

Status
Not open for further replies.
Feb 2, 2008
202
0
0
42
#1
What did they do? Can someone reproduce their method and get the same answer?

Because the normal method that everyone else uses is:

27 FLOPS per cycle per pixel pipe.
24 pixel pipes.
27*24=648
10 FLOPS per cycle per vertex pipe.
8 vertex pipes.
10*8=80
648+80=728 total FLOPS per cycle.
500MHz=500,000,000 cycles per second.
728*500,000,000=364,000,000,000 floating point operations per second=
364 GFLOPS=

0.364 TFLOPS

So yeah, 0.364<1.8

So did I do something wrong? Does Sony know something we dont?

I understand that 1.8 TFLOPS is probably PR BS but what method did they use to get that number?
 

seebs

Elite Sage
Dec 29, 2006
11,822
0
0
50
#2
Well, it DOES have components other than pixel and vertex pipes. But yeah, it does seem a bit off.

Perhaps instead of using the normal method, they used some other method where they can count more operations -- say, counting a fused multiply-add as two or three operations.
 
R

renegadevikingPS3

Guest
#3
1.8 Teraflops floating point for the machine. Anyhow, the actual usable flops is around 440-460 Gigaflops.

Source: Wherever PS3 specs are found.
 
Feb 2, 2008
202
0
0
42
#4
[QUOTE="renegadevikingPS3, post: 0]1.8 Teraflops floating point for the machine. Anyhow, the actual usable flops is around 440-460 Gigaflops.[/quote]

no 1.8 T was stated for rsx itself. 218 G was stated for cell. the system total was stated as 2.18 TFLOPS. ALL stated by sony.
 

Centurion

Dedicated Member
May 15, 2008
1,014
5
0
#5
I came up with 1.8 TFLOPS using a 3-dimensional dot product as 5 operations (3 multiplies and 2 adds)

27 x 24 x 5 = 3240

10 x 8 x 5 = 400

(3240 + 400) x 500,000,000= 1.8 x 10^12 = 1.8 trillion
 
Feb 2, 2008
202
0
0
42
#6
[QUOTE="seebs, post: 0]Well, it DOES have components other than pixel and vertex pipes. But yeah, it does seem a bit off.

Perhaps instead of using the normal method, they used some other method where they can count more operations -- say, counting a fused multiply-add as two or three operations.[/quote]

i see so they may have broken down operations into more, simpler operations so that they could count more.

also do you think that pure video is in there? do they count disabled transistors as part of the transistor count?
 
Feb 2, 2008
202
0
0
42
#7
[QUOTE="Centurion, post: 0]I came up with 1.8 TFLOPS using a 3-dimensional dot product as 5 operations (3 multiplies and 2 adds)

27 x 24 x 5 = 3240

10 x 8 x 5 = 400

(3240 + 400) x 500,000,000= 1.8 x 10^12 = 1.8 trillion[/quote]

oh, well that explains it then! thankyou.:)
 

Centurion

Dedicated Member
May 15, 2008
1,014
5
0
#8
I could also be way off, though. There are too many unknowns. For all we know the floating-point units could run at twice the speed of the GPU. We don't know the instruction executino time or latency, or even what instructions are used to get the 1.8TFLOP figure.
 
P

playstnlvr

Guest
#9
I've seen the math done before. Calculating theoretical peak floating point performance isn't very hard. It's only theoretical though, but then again, so is everything you would compare it to.
 
R

renegadevikingPS3

Guest
#10
Lies. Nvidia said that a x2 G92 Geforce 9800 GX2 does 1+ Teraflop floating point while the PS3 has a G71 with slower memory. Logically it doesn't make sense. No G71 x86 GPU had 1+ Teraflop power.
 
C

Comatosed

Guest
#12
[QUOTE="The Game, post: 0]It does it really matter? Does anyone really care?[/quote]

Somebody must or we wouldn't keep seeing these damn spec threads :( :help
 

seebs

Elite Sage
Dec 29, 2006
11,822
0
0
50
#13
[QUOTE="The Game, post: 0]It does it really matter? Does anyone really care?[/quote]

There's no point in playing games without an occasional argument about specs. It'd be like playing baseball and not talking about the stats of the various players. CRAZY!
 
S

SnowHawk

Guest
#14
like someone said does it really matter and does anyone really care about this stuff?

just enjoy the games that is what matters.
 

TheRocker

Apprentice
Apr 26, 2008
219
0
0
81
#15
[QUOTE="renegadevikingPS3, post: 0]Lies. Nvidia said that a x2 G92 Geforce 9800 GX2 does 1+ Teraflop floating point while the PS3 has a G71 with slower memory. Logically it doesn't make sense. No G71 x86 GPU had 1+ Teraflop power.[/quote]

What exactly is lies? RSX specs are under tight NDA, nobody knows anything about it, if they tell you they do, or "it's like GF7800GTX" or other crap, they are simply spewing internet forum talk, pretending to sound intelligent...
 

Vulgotha

Power Member
Jan 6, 2007
15,776
148
0
31
#16
It's an interesting number. I'm dubious as to how they reached it... But it was something Microsoft was never able to "Dispel" with their own math based calculations for the 360.

So apparently they just ignored it... Though I do recall a 240Gflop count going around for Xenos..?

Not that the Terraflops number really means much of anything in all practicality lol.
 

chugs

Apprentice
Oct 16, 2006
430
0
0
42
#17
[QUOTE="snowhawk, post: 0]like someone said does it really matter and does anyone really care about this stuff?

just enjoy the games that is what matters.[/quote]


^ if you don't like this threads then don't particpate with banal posts. Far out. I don't like socket puppets and yet do you see me in their forums making stupid posts? Yes some people like talking about the specs.

Moving on can a G71 core do 1.8tflops? If Nvidia G90 series is only just breaking 1tflops it would seem someone is fudging the numbers, or worse >insert shocking gasp< -=7=- is right, the RSX is a super-uber chip!
 
Nov 10, 2005
4,420
0
0
#20
I remember Nvidia saying the RSX is something like a 7800GTX with a few additional surprises. Don't ask for a source this was like two years ago.

It was probably crap anyway, just Sony mouthing off. I remember when the PS3 was first revealed Sony claimed it was twice as powerful as the 360.
 
Feb 18, 2008
107
0
0
40
#21
[QUOTE="renegadevikingPS3, post: 0]Lies. Nvidia said that a x2 G92 Geforce 9800 GX2 does 1+ Teraflop floating point while the PS3 has a G71 with slower memory. Logically it doesn't make sense. No G71 x86 GPU had 1+ Teraflop power.[/quote]

EXACTLY!
So lets say RSX has a "THEORETICAL" 1.8 teraflop floating point while a 9800 gx2 has 1+ "THEORETICAL" Flops, then ps3 somehow still has more...Theoreticaly speaking ofcourse :mrgreen:

RSX is a new brand of GPU, not the same as GPU'S inside of PC'S.

What really is weird tho, is that sony has anounced a computing board with the same RSX/cell combo that could produce a 4k image resolution.
I know this has been posted beforeand i (Doubt it would be for games) but its there...:)

It just makes you wonder, although there isnt solid proof, we all know that RSX is nothing like a a g70.
Even developers have stated in other forums, RSX cannot be compared to pc gpu's.

It just dosnt make any sence that sony would spend so much money on state of the art bluray and cell ect...Only to have a slightly modified g70 chip in the system....
Sony are smarter than that trust me.
Im not really any genuis when it comes to computer related things..but you would be blind not to see that theres somthing strange about RSX.
Only time will tell, and only a selected few truly understand this.

--

The "Cell Computing Board" prototype, which is small enough to be mounted into a 1U-size server for a 19-inch rack mount, combines the Cell Broadband Engine microprocessor and RSX graphics processor. This pairing results in a high-power computing platform that goes beyond the power of the Cell chip alone
http://www.infoworld.com/article/07/08/01/Sony-Cell-based-computing-board_1.html

---

The incorporation of RSX realizes arithmetic operation speeds beyond the 230 GFLOPS of the Cell/B.E. microprocessor alone
http://www.sony.net/SonyInfo/News/Press/200708/07-070E/index.html
 

mynd

Ultimate Veteran
May 3, 2006
20,883
191
63
49
Down Under
#22
Funny argument, given half the flops in the vertex shader will not be used.

If GeomEdge is doing it's task properly, that is.
 

seebs

Elite Sage
Dec 29, 2006
11,822
0
0
50
#23
No, highdef, we do not "all know" that RSX is nothing like a G70.

We "all know" you are a liar.

Both Sony and nVidia have claimed that RSX is similar to a G70. They have cited operating characteristics and transistor counts. Developers have commented on the similarities.

Meanwhile, you post outright lies, over and over. Sony says RSX is more powerful than a pair of GeForce 6800s; you claim they said it was more powerful than four 7800s. Why? Because if you quote what they actually said, everything points to it being a close relative of the G70, probably with more texture cache. Nothing points to it being different in any significant way.

Except, of course, your habitual outright lies.

To be fair, in this post, apart from the usual stuff, you simply can't understand basic English. You seem much impressed by the discovery that Cell+RSX has computational power beyond that of Cell alone. How can this be surprising or impressive? That's what always happens when new chips are added -- it's why we use more than one chip in many systems. Multiple chips have, in general, more processing power than any one of the chips on its own. Duh.

Edited:

http://forum.beyond3d.com/showthread.php?t=25392

"The 1.8 TFlops number isn't 'real' in the opinion of many; certainly myself included. It includes a lot of fixed-function nonsense that really doesn't factor in to the programmable ops people normally concern themselves with when deriving these figures."

So what they're really doing is things like claiming 27 flops per pixel pipeline even though that's not really what happens.

http://www.neogaf.com/forum/showthread.php?t=284689

"What you are thinking of, is either the 2.18 TFLOP given for PS3 CPU+GPU (total system) or the 1.8 TFLOPs for just the GPU alone. This is a PR stunt that was used to inflate PS3 spec to beat Xbox 360's equally nonsensical PR spec of 1 TFLOP."

As someone else points out:

"Let's compare the PS3's RSX to a 7600GT!


550mhz core vs 560mhz core.
128 bit memory interface vs 128 bit memory interface.
22.4GBps memory bandwidth vs 22.4GBps memory bandwidth.
4.4 billion pixel fillrate vs 4.4 billion pixel fillrate.
8 ROPs vs 8 ROPs.
256mb GDDR3 vs 256mb GDDR3"

Happy now?

It's a made-up number, same as the ludicrously overstated numbers MS gave for the 360. It's not a real-world number.
 
May 20, 2008
10,998
131
63
#25
[QUOTE="seebs, post: 0]No, highdef, we do not "all know" that RSX is nothing like a G70.

We "all know" you are a liar.

Both Sony and nVidia have claimed that RSX is similar to a G70. They have cited operating characteristics and transistor counts. Developers have commented on the similarities.

Meanwhile, you post outright lies, over and over. Sony says RSX is more powerful than a pair of GeForce 6800s; you claim they said it was more powerful than four 7800s. Why? Because if you quote what they actually said, everything points to it being a close relative of the G70, probably with more texture cache. Nothing points to it being different in any significant way.

Except, of course, your habitual outright lies.

To be fair, in this post, apart from the usual stuff, you simply can't understand basic English. You seem much impressed by the discovery that Cell+RSX has computational power beyond that of Cell alone. How can this be surprising or impressive? That's what always happens when new chips are added -- it's why we use more than one chip in many systems. Multiple chips have, in general, more processing power than any one of the chips on its own. Duh.

Edited:

http://forum.beyond3d.com/showthread.php?t=25392

"The 1.8 TFlops number isn't 'real' in the opinion of many; certainly myself included. It includes a lot of fixed-function nonsense that really doesn't factor in to the programmable ops people normally concern themselves with when deriving these figures."

So what they're really doing is things like claiming 27 flops per pixel pipeline even though that's not really what happens.

http://www.neogaf.com/forum/showthread.php?t=284689

"What you are thinking of, is either the 2.18 TFLOP given for PS3 CPU+GPU (total system) or the 1.8 TFLOPs for just the GPU alone. This is a PR stunt that was used to inflate PS3 spec to beat Xbox 360's equally nonsensical PR spec of 1 TFLOP."

As someone else points out:

"Let's compare the PS3's RSX to a 7600GT!


550mhz core vs 560mhz core.
128 bit memory interface vs 128 bit memory interface.
22.4GBps memory bandwidth vs 22.4GBps memory bandwidth.
4.4 billion pixel fillrate vs 4.4 billion pixel fillrate.
8 ROPs vs 8 ROPs.
256mb GDDR3 vs 256mb GDDR3"

Happy now?

It's a made-up number, same as the ludicrously overstated numbers MS gave for the 360. It's not a real-world number.[/quote]

What ludicrously overstated numbers are you talking about.
 

seebs

Elite Sage
Dec 29, 2006
11,822
0
0
50
#26
[QUOTE="Sub-stance1, post: 0]What ludicrously overstated numbers are you talking about.[/quote]

Normally, when discussing the performance of a GPU, we talk about the number of operations it performs per second -- FLoating point Operations Per Second. FLOPS. These come in millions (MFLOPS), billions (GFLOPS -- and that's "thousand million" to you crazy foreigners), and quadrillions (TFLOPS -- and that's "billion" to you crazy foreigners).

Now, there are two ways for these operations to exist. One is what's called "fixed-program", where there is actual silicon hardwired to perform a particular operation. The other is "programmable", where the silicon is executing instructions given to it. The former is much faster, but only the latter is generally relevant to game programmers.

When Microsoft announced specs for Xenos, they claimed ludicrously high numbers -- around 1TFLOP -- which are far and away beyond the measured performance of the chip. How did they do this? Why, by including all the fixed program operations. So, Sony came out and said the PS3 was two TFLOPS, by which they meant about 230GFLOPS for Cell, plus another 1.8TFLOPS for RSX.

But both are, in effect, lying; they are reporting a number having nothing at all to do with actual performance while rendering games.

Xenos is officially rated at ~240GFLOPS counting only pixel shaders. RSX looks to have more operations per pipeline and fewer pipelines, probably.
 

Centurion

Dedicated Member
May 15, 2008
1,014
5
0
#27
[QUOTE="seebs, post: 0]Normally, when discussing the performance of a GPU, we talk about the number of operations it performs per second -- FLoating point Operations Per Second. FLOPS. These come in millions (MFLOPS), billions (GFLOPS -- and that's "thousand million" to you crazy foreigners), and quadrillions (TFLOPS -- and that's "billion" to you crazy foreigners).
[/quote]

A TFLOP is a trillion floating point operations per second.
 

seebs

Elite Sage
Dec 29, 2006
11,822
0
0
50
#28
[QUOTE="Centurion, post: 0]A TFLOP is a trillion floating point operations per second.[/quote]

DOH.

You are, of course, right.

Okay. So, change "quadrillion" back to "trillion" (I think I actually got the British usage right) and it'll make sense. 1 tflop = 1000 gflop = 1,000,000 mflop = 1,000,000,000,000 flop.

Which is four sets of three zeroes, which is why I thought of "quad". Sorry!
 

surtur

Super Elite
Oct 5, 2005
2,134
0
0
40
#29
Ok one thing i need to mention is the time that those specs came out anyone remember ??? i am pretty sure it was while Nvidia was still making the chips and all those specs were just on paper and what it should be able to do in theory. and since the real numbers and anything else about RSX is under top secret classified armed guard in Area 51 (you know they did get the tech from those darn alien friends of ours right lol) they have no reason to release updated specs esp if it doesn't live up to what they said it would.

As for do i care nope i am enjoying my gaming and it doesn't make a difference to me if the PS3 was 100x more powerful then it is now i am happy with what i got.
and to all you guys that every time a RSX thread comes out and start getting all hissy with each other you are all wrong until the real specs and numbers are released if ever. all you are doing is repeating yourselves with half truths and a whole bunch of rumor's trying to sound more technical then the last guy so you sound more believable yet you know just as much about the darn thing as anyone else that has seen the same specs you have.
I am not saying you guys don't know what you are talking about (in a technical sense) but unless you have actually worked with or were one of the people designing the chip you have nothing to back up what you say.
 

chugs

Apprentice
Oct 16, 2006
430
0
0
42
#30
[QUOTE="seebs, post: 0]No, highdef, we do not "all know" that RSX is nothing like a G70.[/quote]

hehe its like the cold war with you two.

[QUOTE="seebs, post: 0]"What you are thinking of, is either the 2.18 TFLOP given for PS3 CPU+GPU (total system) or the 1.8 TFLOPs for just the GPU alone. This is a PR stunt that was used to inflate PS3 spec to beat Xbox 360's equally nonsensical PR spec of 1 TFLOP."[/quote]

Irrespective of whether its a PR stunt, irrespective of whether the above values are meaningful for developers and irrespective of whether the values indicate a meaningful render, could a G71 GPU perform 1,500,000,000,000 floating point calculations?
 
Status
Not open for further replies.