Support NeoGAF

[Kleegamefan]

Could it be STI saw the importance of integer math performance and gave it a bump in the DD2 rev of CELL for PS3...

Check it:

http://www-128.ibm.com/developerworks/power/library/pa-cellperf/

The SPUs SIMD support can perform operations on sixteen 8-bit integers, eight 16-bit integers, four 32-bit integers, or four single-precision floating-point numbers per cycle. At 3.2GHz, each SPU is capable of performing up to 51.2 billion 8-bit integer operations or 25.6GFLOPs in single precision.

This is from IBMs newest PS3 CELL doc dated Nov 22, 2005


Back at ISSCC 2005, the DD2 version of a CELL could do the same number of Gigaflops per SPU as integer ops.....

On a 3.2Ghz PS3 version of CELL, that would be 25.6B floating point operations per sec and 25.6B integer ops per second(per SPU, multiplied by 7 for the entire CELL CPU)....but now we have clarification that an SPU can work on 16, 8-bit integers at the same time and the number now jumps up to 51.2B integer ops per sec.....which is a number double that of the quoted SPU floating ops performance number (25.6Gflops).....

So that would mean that, independent of the PPE, ,the 7 CELL SPUs in PS3 can do 179.2Gflops and 358.4 Billion 8-bit integer ops per second!!!!

As a point of reference, a 3.06 GHz Pentium 4 has a theoretical peak of 12 gigaflops!!!

Perhaps the original 25.6B integer op figures were for 16-bit integers? (of which, an SPU can work on 8 at a time!!)

This thing is a beast...

 

[Amir0x]

I just can't wait to see how them big Japanese devs exploit this bad boy. Metal Gear Solid 4 in all its glory... man. Gonna be some day. The game is gonna have bystanders too. Of course I'm going to commit genocide in the game so that country A, B and C all hate me.

 

[The Take Out Bandit]

It is the key to winning the console war! :lol

 

[xexex]

in before Panajev - come on Pana help us out here

 

[GhaleonEB]

Could it be STI saw the importance of integer math performance and gave it a bump in the DD2 rev of CELL for PS3...

Check it:

http://www-128.ibm.com/developerworks/power/library/pa-cellperf/




This is from IBMs newest PS3 CELL doc dated Nov 22, 2005


Back at ISSCC 2005, the DD2 version of a CELL could do the same number of Gigaflops per SPU as integer ops.....

On a 3.2Ghz PS3 version of CELL, that would be 25.6B floating point operations per sec and 25.6B integer ops per second....but now we have clarification that an SPU can work on 16, 8-bit integers at the same time and the number now jumps up to 51.2B integer ops per sec.....which is a number double that of the quoted SPU floating ops performance number (25.6Gflops).....

So that would mean that, independent of the PPE, ,the 7 CELL SPUs in PS3 can do 179.2Gflops and 358.4 Billion 8-bit integer ops per second!!!!

As a point of reference, a 3.06 GHz Pentium 4 has a theoretical peak of 12 gigaflops!!!

Perhaps the original 25.6B integer op figures were for 16-bit integers? (of which, an SPU can work on 8 at a time!!)

This thing is a beast...

Don't hold back. Tell us how you really feel.

 

[TTP]

in before gofreak

 

[fse]

it wont have enough vents.

 

[DonasaurusRex]

..well 8 bit integers are only numbers ranging from 0 - 255 and since they are integers they dont have to worry about precision or numbers computers have a little difficulty handling due to memory being finite but numbers involving precision can be infinite like .3333... . Im guessing each SPU is 128 bit capable so of course the integer performance is going to be off the charts with multiple spu's and the ability to handle that many 8bit data chunks per cycle.

if the rate of the cell is 3.2 billion clock pulses per cycle, and it can handle 16 8 bit integers per cycle..then you get (16 8 bit data chunks x 3.2B pulses per cycle )= 51.2 billion ops per second per SPU. You up that to 16 bits and its that same number you quoted for floating point numbers. So im guessing regular floating point numbers on teh cell are 16bits to the cpu, how that is ordered internally how many bits you get for the number and precision i dunno. However Gflops refer to floating point operations , unless that has changed so I think thats why the Gflops number reported is so much lower than the 8 bit integer number reported because of the precision needed in floating points they include precision bits so i guess IBM decided to use 16 bits for floats.

 

[Date of Lies]

it's actually 0 to 127 if they're signed integers, which they most likely are.

 

[DonasaurusRex]

yes you are right forgot about that, 0-255 is representable, 256 actual numbers, divide by two for signed ints. Im not sure if signed or unsigned ints are the norm for benchmarking theoretical peaks.

 

[Mrbob]

PS3 ftw?

 

[Red Scarlet]

Ya got me, klee.

 

[ToxicAdam]

 

[Goreomedy]

That's the most pitiful graphic I've ever seen.

And because of that, I laughed.

 

[BlueTsunami]

 

[Shogmaster]

Could it be STI saw the importance of integer math performance and gave it a bump in the DD2 rev of CELL for PS3...

Check it:

http://www-128.ibm.com/developerworks/power/library/pa-cellperf/




This is from IBMs newest PS3 CELL doc dated Nov 22, 2005


Back at ISSCC 2005, the DD2 version of a CELL could do the same number of Gigaflops per SPU as integer ops.....

On a 3.2Ghz PS3 version of CELL, that would be 25.6B floating point operations per sec and 25.6B integer ops per second(per SPU, multiplied by 7 for the entire CELL CPU)....but now we have clarification that an SPU can work on 16, 8-bit integers at the same time and the number now jumps up to 51.2B integer ops per sec.....which is a number double that of the quoted SPU floating ops performance number (25.6Gflops).....

So that would mean that, independent of the PPE, ,the 7 CELL SPUs in PS3 can do 179.2Gflops and 358.4 Billion 8-bit integer ops per second!!!!

As a point of reference, a 3.06 GHz Pentium 4 has a theoretical peak of 12 gigaflops!!!

Perhaps the original 25.6B integer op figures were for 16-bit integers? (of which, an SPU can work on 8 at a time!!)

This thing is a beast...

OMG THAT SOUNDS FUCKING INCREDIBLE!! *goes back to playing with X360*

 

[BlueTsunami]

*goes back to playing with X360*

:_;

I'll be getting mine soon....I haven't preodered anything..I'm going to wander the streets looking for it...but I will find one...

 

[Bigfoot]

it's actually 0 to 127 if they're signed integers, which they most likely are.

Actually, it's -128 to 127, which isn't really that much different than 0 to 255, one is signed and one isn't, but the basic operations are still the same.

 

[Doube D]

OMG THAT SOUNDS FUCKING INCREDIBLE!! *goes back to playing with X360*

no no shogs, the first step is admission. You are still in denial... sniff

 

[aaaaa0]

Vector integer math operations are nothing new. 10+ years ago Pentium had MMX, which is the same idea.

The important issues for integer performance are load/store, conditionals and branching.

 

[∀ Narayan]

LAUGHTER!

 

[hellfire]

Vector integer math operations are nothing new. 10+ years ago Pentium had MMX, which is the same idea.

The important issues for integer performance are load/store, conditionals and branching.

yup. altivec has always supported these vector integer operations. and anyway, it won't help general purpose computing.

 

[kaching]

I just can't wait to see how them big Japanese devs exploit this bad boy. Metal Gear Solid 4 in all its glory... man. Gonna be some day. The game is gonna have bystanders too. Of course I'm going to commit genocide in the game so that country A, B and C all hate me.

Don't forget to take a little time out from all that warmongering to trim your 'stache. Hygiene is very important on the battlefield.

 

[Vustadumas]

Yep, mass integer calculations is what makes games great! PS3 FTW!

 

[Mefisutoferesu]

Yep, mass integer calculations is what makes games great! PS3 FTW!

?? Sarcasm??

Obviously, there isn't any one thing that makes games great, but I imagine you're talking in a "power" sorta kinda way; getting pretty visuals and fancy ani/phys, right? That's more a product of Floating Point ops than integer. While the majority of code written in games does tend to be integer based I'd say the (vast?) majority of execution time is spent in floating point operations. Well, it might be a slightly different story for some odd game here or there, but it applies pretty well across the board.

 

[gofreak]

Could it be STI saw the importance of integer math performance and gave it a bump in the DD2 rev of CELL for PS3...

Check it:

http://www-128.ibm.com/developerworks/power/library/pa-cellperf/


This is from IBMs newest PS3 CELL doc dated Nov 22, 2005


Back at ISSCC 2005, the DD2 version of a CELL could do the same number of Gigaflops per SPU as integer ops.....

On a 3.2Ghz PS3 version of CELL, that would be 25.6B floating point operations per sec and 25.6B integer ops per second(per SPU, multiplied by 7 for the entire CELL CPU)....but now we have clarification that an SPU can work on 16, 8-bit integers at the same time and the number now jumps up to 51.2B integer ops per sec.....which is a number double that of the quoted SPU floating ops performance number (25.6Gflops).....

So that would mean that, independent of the PPE, ,the 7 CELL SPUs in PS3 can do 179.2Gflops and 358.4 Billion 8-bit integer ops per second!!!!

As a point of reference, a 3.06 GHz Pentium 4 has a theoretical peak of 12 gigaflops!!!

Perhaps the original 25.6B integer op figures were for 16-bit integers? (of which, an SPU can work on 8 at a time!!)

This thing is a beast...

The 25.6B integer ops figure would be for 32-bit integers.

Simply have a figure for integer ops only tells us so much. When you dig deeper and notice, for example, that certain operations aren't available for certain integer types, and you see that latencies are different it becomes clearer that the SPUs are indeed optimised for FP. That's how it should be

 

[Amir0x]

Don't forget to take a little time out from all that warmongering to trim your 'stache. Hygiene is very important on the battlefield.

As long as it showcases THE POWER OF PS3™.

 

[Christoff Yurievich]

hang on, isn't this the same discussion, 15 hours earlier?

So, does this mean that the "tech demos" are probably running on the actual processor and not rendered? I know MGS4 was touted as being run on a cell chip, but Warhawk and the like (as far as I know) were still in that iffy space where it could be either cell or rendered.

 

[gofreak]

So, does this mean that the "tech demos" are probably running on the actual processor and not rendered? I know MGS4 was touted as being run on a cell chip, but Warhawk and the like (as far as I know) were still in that iffy space where it could be either cell or rendered.

Warhawk was certainly realtime on Cell (+"RSX" of course).

Cell dev kits have been available since early this year. Most or any realtime stuff we've seen for PS3 has been running with Cell.

 

[Brofist]

OMG THAT SOUNDS FUCKING INCREDIBLE!! *goes back to playing with X360*

good idea that you get some time in with it before it's obsolete..

i kid

 

[Panajev2001a]

in before Panajev - come on Pana help us out here

What's there to help... even the EE's MMI (the 128-bits SIMD instructions of the twin 64 bits ALU's inside the EE) was capable of up to 16 operations on 8 bits integers.

While we use FP with 32 and 64 btis precision mostly (single and double precision FP), we use chars (8 bits signed or unsigned), 16 bits integers, 32 bits integers, 64 bits integers, etc...

Ina single 128 bits register you can hold:

1x128 bits value

2x64 bits values

4x32 bits values

8x16 bits values

16x8 bits values


No magic .

 

[Panajev2001a]

The 25.6B integer ops figure would be for 32-bit integers.

Simply have a figure for integer ops only tells us so much. When you dig deeper and notice, for example, that certain operations aren't available for certain integer types, and you see that latencies are different it becomes clearer that the SPUs are indeed optimised for FP. That's how it should be

Someone did forget the need of directly multiplying 32 bits integers .

 

[Shadow Moses]

So that would mean that, independent of the PPE, ,the 7 CELL SPUs in PS3 can do 179.2Gflops and 358.4 Billion 8-bit integer ops per second!!!!

As a point of reference, a 3.06 GHz Pentium 4 has a theoretical peak of 12 gigaflops!!!

Perhaps the original 25.6B integer op figures were for 16-bit integers? (of which, an SPU can work on 8 at a time!!)

This thing is a beast...

OMGWTFBBQ

Cell is ike 14.933333333333333 times more powerful than Pentium 4!!!!!!!!!!

We better put some safety measures on PS3 release, as Cell can me used to manufacture Nuclear weapon and thereby endorsing terrorism!!111!

OMG PS3 supports terrorism! !!!1111111111!! :lol




Never seem such a pointless PR shit in my entire life :\

 

[BlueTsunami]

OMGWTFBBQ

Cell is ike 14.933333333333333 times more powerful than Pentium 4!!!!!!!!!!

We better put some safety measures on PS3 release, as Cell can me used to manufacture Nuclear weapon and thereby endorsing terrorism!!111!

OMG PS3 supports terrorism! !!!1111111111!! :lol




Never seem such a pointless PR shit in my entire life :\

...

 

[teiresias]

OMGWTFBBQ

Cell is ike 14.933333333333333 times more powerful than Pentium 4!!!!!!!!!!

We better put some safety measures on PS3 release, as Cell can me used to manufacture Nuclear weapon and thereby endorsing terrorism!!111!

OMG PS3 supports terrorism! !!!1111111111!! :lol




Never seem such a pointless PR shit in my entire life :\

I hate morons that don't know the difference between white papers and PR. I'd hate to see some of you do searches through the IEEExplore literature database, "OMG!! This paper published in the IEEE Consumer Electronics Society Proceedings is such a pointless piece of PR!"