Support NeoGAF

[midnightguy]

http://xbox360.ign.com/articles/612/612995p1.html

the GPU, the Graphic Processing Unit, which handles the graphic output of the system (and which has 150 million transistors in it), the Southbridge, which enables all of the Ethernet and controller issues, and the TV encoder, which handles resolution issues, such as progressive scan, interlacing and other TV related issues.

The GPU has the ability to generate 48 shader units using an advanced shader language. In the standard PC, you have both vertex shaders and pixel shaders, each working individually. In the Xbox 360, the Shader units can do either, meaning you can have scenes with only a little bit of geometry and a tremendous number of effects, such as a fightng game with only two characters on screen (like Dead Or Alive 4). Or you can deliver tons of geometry with fewers effects, such as a a 100-person action online game (such as Possession or Huxley). It has 512 MB of main memory with a 10MB framebuffer. And it's got 10MB embedded DRAM.

that seems awfully low.

the old Radeon 9700 had 107 million, the 9800 had 110 million, the X800 has 160 million and the upcoming R520 Fudo is meant to have at least 300 million.

maybe this explains the "meh" realtime in-game graphics of many games (the completely awesome looking stuff seems to be pre-rendered CG) and also why Xbox360 only has a fillrate of 4 gigapixels (with 4x AA yes i know), and can only produce 500 million polygons/sec. the X800 can produce over 8 gigapixels and 600 million polygons/sec.

the 48 unified shader "pipes" are really just ALUs, and there are 2, 3 or 4 ALUs to a typical pipeline in PC GPUs.

it seems that Xbox360 GPU might only be the equivalent of 12 pipelines total, in terms of raw performance.

or maybe this is way off.


if the GPU is only 150 million transistors, that is slightly less than 2.5 times more than Xbox GPU which was like 63 million transistors.

 

[Oogami]

Maybe this is why Team Ninja not happy with it and they're the great defection? They like POWER.

+1 PS3

 

[Izzy]

Just for the reference - my NV40's got 220 million transistors.

And new PS3 GPU is supposedly >>>> NV40 SLI set-up :lol

 

[CabbageRed]

Going by Gears of War, I am wondering how much I should care about transistors when it comes to actual games on the Xbox 360.

 

[Monk]

How much does transistor count affect graphical power?

 

[duckroll]

Perfect Dark Zero's character models - 5000 polys

Perfect Dark Zero's enviroment models - 100 thousand polys

Xbox360's hardware GPU - 150 million transistors

The look of Bill Gates' face when he read gaming forums right after the MTV event - priceless

:lol :lol :lol

 

[j^aws]

I'll just cut and paste what I just posted on Beyond3D,

CPU and GPU transistor counts shouldn't be comapred. Heat dissipation is directly proportional to transistor count and clock frequency as a general rule. GPUs ~ 500 MHz and CPUs ~ 3 GHz. And you shouldn't definitely be comparing chips on different processes.

Comapre CPU to CPU transistor counts on the SAME process, e.g. 90nm.

 

[Jonnyram]

How many transistors does real life use to gain its photographic quality?

 

[midnightguy]

this cannot be right though. 150 million transistors? if the CPU is 165 million, the GPU has less. that has not been the case with game consoles the last few generations. the GPUs are always larger and have more trannies than the CPU.

 

[Vince]

How much does transistor count affect graphical power?

Generally they scale quite well when you look at it more abstractly, it's not a perfect 1:1 mapping, but they are intimately related as the more area devoted to logic you have, the more computation you can preform, etc.

In all honestly, as much as I'd like to believe this, I find it hard to believe that a 90nm - basically clean-sheet design- design has only 130M transistors. Especially for one that is as shader|computation focused as the R500.

The only rational way I can see this happening is if the 130M are purely logic transistors and they had to trade-off logic area for that 10MB of eDRAM. If this isn't the case, then there's something very wrong.

How many transistors does real life use to gain its photographic quality?

I know you're being a smartass, and I dont mean to get in your way, but this is something I know quite a few physicist are interested in... granted the question is more about information and matter and studying this viz-a-viz black-hole entropy loss, but it's the same thing in spirit.

 

[Ryudo]

this cannot be right though. 150 million transistors? if the CPU is 165 million, the GPU has less. that has not been the case with game consoles the last few generations. the GPUs are always larger and have more trannies than the CPU.

I think he means the die size, like 60nm or 90nm etc. A die shrink would (should) equal less heat and less transistors.

 

[j^aws]

this cannot be right though. 150 million transistors? if the CPU is 165 million, the GPU has less. that has not been the case with game consoles the last few generations. the GPUs are always larger and have more trannies than the CPU.

WTF are you talking about???

You guys are clueless...

More transistors means more heat...

Higher clock means more heat...

Smaller process means more transistors...

PHYSICS plays it's part...

Just do a general PS2 comaprisosn,

i.e. EE ~ 13 mil @ 300 MHz, 250nm

and GS ~ 40 mil @ 150 MHz, 180nm

I repeat, APPLES to APPLES. If anything compare X360 to CELL at 90nm...but stop with this random BS!

 

[Vince]

WTF are you talking about???

You guys are clueless...

Hey douche, chill. Lets look at who is clueless:

You give two premises:

1: More transistors means more heat...
2: Higher clock means more heat...

Lets see how this works in reality using your EE and GS (both 250nm designs) example concerning their logic counts and the effects your premises have upon them:

Emotion Engine, ~12M logic tranistors @ 300MHz
Graphic Synthesizer, ~7M logic transistors @ 150MHz

Wow, your premise and example went far....


Now, his statement of GPUs having higher logic densities (as compared to CPUs) recently is grounded in truth as modern GPUs follow the stream processing paradigm and, thus, have lesser need for cache, while biasing towards computation. This means that the designs have less SRAM and more logic, logic being more dense and, thus, the design having moe logic gates and higher transistor counter per mean area than a CPU.

 

[Sqrt_minus_one]

I think he means the die size, like 60nm or 90nm etc. A die shrink would (should) equal less heat and less transistors.

Less transistors ? Surely you mean area.

WRT the B3D comment what process are the CPU and GPU made on respectively I thought both were 90nm.

 

[MbGCaM]

Transistors =/= Power.

 

[Raven.]

Generally they scale quite well when you look at it more abstractly, it's not a perfect 1:1 mapping, but they are intimately related as the more area devoted to logic you have, the more computation you can preform, etc.

In all honestly, as much as I'd like to believe this, I find it hard to believe that a 90nm - basically clean-sheet design- design has only 130M transistors. Especially for one that is as shader|computation focused as the R500.

The only rational way I can see this happening is if the 130M are purely logic transistors and they had to trade-off logic area for that 10MB of eDRAM. If this isn't the case, then there's something very wrong.

Lower cost- aka cheaper to produce. But that's gonna bite'em big time if sony decides to push the envelope(the ps2 was supposed to have far more ram, aka probably 64 and maybe even 128... significant yield problems kept sony from being more ambitious at that time(8MBedram, 128MB main ram))... this time though with ibm/toshiba and having perfected the process, and reporting good yield... I wouldn't put it pass them to throw a 2-cell setup(after all projections in many magazine and engineering sites were for a cpu with 500M at .1micron several years ago) with an uber gpu(500~M transistors)....

 

[midnightguy]

my whole point is, Xenon / 360 GPU seems to be awfully underpowered for a true next-gen console. it seems to be a smaller leap than any other move to next-generation. I hope to be completely wrong though. and I also hope that any CGI that we know of, and any other games that were CGI, that looked great, can be done real-time on Xbox360.

 

[Nostromo]

and which has 150 million transistors in it

No way, they're wrong. No doubt about that, 10 mb of edram alone should take 80+ Mtransistors.
R500 is not underpowered! c'mon guys, just read the specs..

 

[Vince]

No way, they're wrong. No doubt about that, 10 mb of edram alone should take 80+ Mtransistors.

I agree as I already stated, Unless, to play devil's advocate, as I threw out there they are talking in just terms of logic. Perhaps the eDRAM, being as area ineffecient as it is compared to pure logic, is doing to the R500 what happened to the GS: 270mm2 of area containing only ~48M transistors and only ~7M of those logic.

 

[midnightguy]

Xbox 360 transistor count so far: 165 + 150 = 315million.

PS3 original transistor count - CPU: 500m. GPU: probably more than 500 million = 1 billion or more transistors.

even if PS3 is now only 500-700 million transistors total (say CPU 250m + GPU 400m) that'd still be a helluva lot more than Xbox360.


of course, work that can be done PER transistor also counts.

well we'll see in a few days

 

[midnightguy]

I agree as I already stated, Unless, to play devil's advocate, as I threw out there they are talking in just terms of logic. Perhaps the eDRAM, being as area ineffecient as it is compared to pure logic, is doing to the R500 what happened to the GS: 270mm2 of area containing only ~48M transistors and only ~7M of those logic.

now THAT would be awesome. If R500s eDRAM to logic transistor ratio was similar to that of Graphics Synthesizer. 6:1 or 4:1 or whatever it actually is.


Playstation2, dispite its crippled amount of RAM, can still do some VERY nice graphics that even Gamecube and Xbox have not managed.

 

[bitwise]

the graphics are great. "ho-hum" depends on which game you were looking at. in which case, the fault lies with the developer and not the hardware. i mean surely you know this.

 

[gofreak]

CPU and GPU transistor counts shouldn't be comapred

Who compared it to CPUs? I've seen other GPU transistor counts mentioned, but no CPU counts.

How much does transistor count affect graphical power?

All else being equal, more transistors should equate to more power.

 

[sol5377]

MY CONSOLE HAS MORE TRANSISTORS THAN YOURS SO THERE!

 

[Nostromo]

Well..everything is possible but 70Mtransistors devoted to logic is just too little.
C'mon..NV2A has almost 70 Mtransistors 4 years ago :lol
They just misquoted their source, I hope the real transistor count will be disclosed soon.
What happened with PS2 GS can't happen (or shouldn't happen) this time, this time is all about shading power, not fillrate, you know.

 

[gofreak]

Well..everything is possible but 70Mtransistors devoted to logic is just too little.
C'mon..NV2A has almost 70 Mtransistors 4 years ago :lol
They just misquoted their source, I hope the real transistor count will be disclosed soon.
What happened with PS2 GS can't happen (or shouldn't happen) this time, this time is all about shading power, not fillrate, you know.

The eDram is in its own module, I doubt they're counting that.

150m ex eDram still seems low, though. The PC GPUs don't have eDram.

 

[midnightguy]

Well..everything is possible but 70Mtransistors devoted to logic is just too little.
C'mon..NV2A has almost 70 Mtransistors 4 years ago :lol
They just misquoted their source, I hope the real transistor count will be disclosed soon.

with this portion of your post, I agree strongly

 

[j^aws]

Hey douche, chill. Lets look at who is clueless:

You give two premises:

1: More transistors means more heat...
2: Higher clock means more heat...

Lets see how this works in reality using your EE and GS (both 250nm designs) example concerning their logic counts and the effects your premises have upon them:

Emotion Engine, ~12M logic tranistors @ 300MHz
Graphic Synthesizer, ~7M logic transistors @ 150MHz

Wow, your premise and example went far....


Now, his statement of GPUs having higher logic densities (as compared to CPUs) recently is grounded in truth as modern GPUs follow the stream processing paradigm and, thus, have lesser need for cache, while biasing towards computation. This means that the designs have less SRAM and more logic, logic being more dense and, thus, the design having moe logic gates and higher transistor counter per mean area than a CPU.

A transistor is a transistor.

You can differentiate these if you want to make a different argument Vince. However my point still stands.

...compare APPLES to APPLES on the SAME process...

Nice to see you miss points here aswell as B3D, meh!

Btw, EE launched at 250nm and the GS launched at 180nm for PS2. If you care to dispute this there's a thread at B3D...

 

[Dr_Cogent]

Non tech guys should never comment on transistor counts, and what not, when they don't know what the fuck they are talking about.

If you aren't a computer engineer, electrical engineer - or have the equivalent knowledge, please - STFU.

 

[Vince]

A transistor is a transistor.

No it's not. There are different physical structures based on switching speed, area, power, etc, and this is just within logic. Look at Cell and it's custom logic, the same was done in key areas of the Pentium4 and all HPreformance processors. Most eDRAM processes have differential structures based on a customers desire for preformance or area; CMOS4 and NEC both offer both.

You can differentiate these if you want to make a different argument Vince. However my point still stands.

Or you can differentiate if you have a clue.

PS. The Graphic Synthesizer was built on a 250nm CMOS process, which is what we need to look at if we wish to debate the process technology, not the economic feasibility:

Microelectronics for Home Applications: Yoshiaki Hagiwara; Sony Corpoartion, Tokyo Japan Fig.18-22, Pg.7-8

 

[Virgin Wii]

my whole point is, Xenon / 360 GPU seems to be awfully underpowered for a true next-gen console. it seems to be a smaller leap than any other move to next-generation. I hope to be completely wrong though. and I also hope that any CGI that we know of, and any other games that were CGI, that looked great, can be done real-time on Xbox360.

I don't know... Test Drive and that Unreal engine game seem to be pretty massive leaps, or at least, IMO they are. About on par with the PSX/Dreamcast transistion.

 

[j^aws]

Who compared it to CPUs? I've seen other GPU transistor counts mentioned, but no CPU counts.

It was posted to stop this random transistor count comparison without knowing what you're comparing.

 

[midnightguy]

hold the phone! according to teamxbox (a totally unbiased source, without a doubt, lol) says that Xbox 360 can process 1500 million aka 1.5 BILLION vertices per second. that makes sense if the polygon performance is 500 million (3 vertices to make a triangle polygon)


http://features.teamxbox.com/xbox/1145/The-Xbox-360-Dissected/p6/

The polygon performance of the Xbox 360 is as high as 500 million triangles per second, which means the Xbox 360 can process some 1.5 billion vertices per second.

the conflicting numbers are giving me a fucking headache :lol

 

[gofreak]

hold the phone! according to teamxbox (a totally unbiased source, without a doubt, lol) says that Xbox 360 can process 1500 million aka 1.5 BILLION vertices per second. that makes sense if the polygon performance is 500 million (3 vertices to make a triangle polygon)


http://features.teamxbox.com/xbox/1145/The-Xbox-360-Dissected/p6/



the conflicting numbers are giving me a fucking headache :lol

Ack. 500m polys != 1.5bn vertices. Vertices are shared between triangles. AFAIK, it's stuck at 500m vertices, but because of sharing of vertices that effectively boils down to ~500m triangles.

 

[j^aws]

No it's not. There are different physical structures based on switching speed, area, power, etc, and this is just within logic. Look at Cell and it's custom logic, the same was done in key areas of the Pentium4 and all HPreformance processors. Most eDRAM processes have differential structures based on a customers desire for preformance or area; CMOS4 and NEC both offer both.

You're STILL missing my point. I've already made my point, APPLES to APPLES. And yes, I know there are different transistors in this world...they ALL give off heat...

Or you can differentiate if you have a clue.

See above. Stop making up arguments to argue with yourself...

PS. The Graphic Synthesizer was built on a 250nm CMOS process, which is what we need to look at if we wish to debate the process technology, not the economic feasibility:

Microelectronics for Home Applications: Yoshiaki Hagiwara; Sony Corpoartion, Tokyo Japan Fig.18-22, Pg.7-8

Yes, I know thank you very much, but I was referring to the launch PS2. I've already made my point APPLES to APPLES...

 

[Nostromo]

Umh..It's not so hard to understand:
Unified ALUs can process vertices and/or pixels.
Let's say all ALUs are processing vertices with a very simple shader (8 instructions).
There are 48 ALUs, this mean every clock cycle it can transform 48/8 = 6 vertices.
6 vertices x 500 Mhz = 3 Bilion vertices per second
(let's give half ALUs to pixel shading and the other half to vcertex shading -> 1.5 Bilion triangles per second)
This is (theoretically) possible.
Too bad the setup engine (a unit that collects transformed vertices and assembles primitives such as triangles) CAN'T setup more than one primitive per clock cycle, so the max triangles per second figure is 500 milions!
Even if the hw can process WAY MORE than 500 milion vertices per second, the setup engine CAN'T setup enough primitives to sustain that rate -> ALUs would stall once internal buffers are full.

 

[Vince]

You're STILL missing my point. I've already made my point, APPLES to APPLES. And yes, I know there are different transistors in this world...they ALL give off heat...

Unless *Shock and Awe* they give off different amounts of heat. I mean who would ever think that a logic gate switching at 150MHz would give off more heat than an eDRAM cell? Or that 13M logic gates switching at 300MHz give off more heat than 7M logic at 150MHz... which really supports your argument *roll*.

To quote you: you're a moron. Your argument is dumb, it's been shown wrong. Stop repeating "Apples to Apples" like it means anything but your attempt at escaping your bullshit responce and attack on that guy. A CPU and a GPU are made of the same fundimental logical operations computed on the same logic, they aren't seperate entities with their own properties just because somone arbitrarily calls one a CPU and one a GPU.

And if you were really thinking about comparing them based off heat, you'd look at the process capability, not what it launched at due to economic reasons. You know, 250nm GS were fabricated prior to nagasaki coming online.... but you just dont get it.

 

[j^aws]

Unless *Shock and Awe* they give off different amounts of heat. I mean who would ever think that a logic gate switching at 150MHz would give off more heat than an eDRAM cell? Or that 13M logic gates switching at 300MHz give off more heat than 7M logic at 150MHz... which really supports your argument *roll*.

To quote you: you're a moron. Your argument is dumb, it's been shown wrong. Stop repeating "Apples to Apples" like it means anything but your attempt at escaping your bullshit responce and attack on that guy.

And if you were really thinking about comparing them based off heat, you'd look at the process capability, not what it launched at due to economic reasons. You know, 250nm GS were fabricated prior to nagasaki coming online.... but you just dont get it.

The point of an argument is to make a point. And I've done that thank you.

You've called me a douche, clueless and a moron in this thread. Explains why you don't post at Beyond3D anymore...See ya...

 

[Vince]

I believe only the douche comment was origional, the others being not mine. And what point did you make? Must have missed it... Then again, you've never been one to think outside of regurgitating what we'd say over and over and posting exoteric comments or simple math to get some useless FP numbers and deductions ad nauseum.

 

[Lazy8s]

Divus Masterei:

the ps2 was supposed to have far more ram, aka probably 64 and maybe even 128... significant yield problems kept sony from being more ambitious at that time(8MBedram, 128MB main ram

It wasn't yield that limited PS2's memory space; it was the system's bandwidth efficiency which dictated that only expensive RAM could be used.

 

[Panajev2001a]

No way, they're wrong. No doubt about that, 10 mb of edram alone should take 80+ Mtransistors.
R500 is not underpowered! c'mon guys, just read the specs..

Was not the e-DRAM off-chip ? ROP's and the rest of AA/Rasterizing logic embedded with DRAM and connected to the GPU chip through a fast bus ?

 

[Shogmaster]

This is the same article that says X360 has "triple core of 3.2 GB processors"? :lol

 

[mrklaw]

Umh..It's not so hard to understand:
Unified ALUs can process vertices and/or pixels.
Let's say all ALUs are processing vertices with a very simple shader (8 instructions).
There are 48 ALUs, this mean every clock cycle it can transform 48/8 = 6 vertices.
6 vertices x 500 Mhz = 3 Bilion vertices per second
(let's give half ALUs to pixel shading and the other half to vcertex shading -> 1.5 Bilion triangles per second)
This is (theoretically) possible.
Too bad the setup engine (a unit that collects transformed vertices and assembles primitives such as triangles) CAN'T setup more than one primitive per clock cycle, so the max triangles per second figure is 500 milions!
Even if the hw can process WAY MORE than 500 milion vertices per second, the setup engine CAN'T setup enough primitives to sustain that rate -> ALUs would stall once internal buffers are full.

nice sums Nostromo. Can you work it backwards for me? i.e. given that the 360 is limited to 500 million triangles per second by setup, how much power does that leave for shaders per pixel? Assume whatever pixel-size you like for triangles that makes sense.

 

[Phoenix]

This is the same article that says X360 has "triple core of 3.2 GB processors"? :lol

Hey at least its no longer (3) 3 Ghz processors
:lol

 

[GhaleonEB]

 

[Lord Error]

R500 is not underpowered! c'mon guys, just read the specs..

I don't get it. According to the original poster, the R500 in Xbox renders somewhat less polys that X800, and has less fillrate, and even technically less pipes. Is that incorrect or is just not directly comparable?

 

[COCKLES]

(pin dropping sound)

 

[Lord Error]

(pin dropping sound)

It looks great, no doubt. But I think anyone who doesn't see that this could be much better looking still (people crowd, flat looking building), is fooling himself :\

 

[Xenon]

It looks great, no doubt. But I think anyone who doesn't see that this could be much better looking still (people crowd, flat looking building), is fooling himself :\

Your right! You can't see into the windows of the buildings. There is no bumped mapped pigeon shit on the statue. There is no gum on the ground, surely someone had to drop gum.

 

[monchi-kun]

damn, that's 1 million short of greatness