G71 has 32 pipelines, 32 pixel shaders & 16 ROPs

[AthlonXP1800]

http://www.theinquirer.net/?article=28609

It going to be very interesting time in next few months, let see the G71 and R580 specs:

GPU Core Memory Pipelines Pixel Shaders ROPs
G71 750 1900 32 32 16
R580 695 1550 16 48 16

It seem Nvidia are doing brilliant job on G71, they dont have problem with 32 pixel shaders into expanded 32 pipelines so ATI are likely to have problems with 48 fat pixel shaders clogged in a thin 16 pipelines, never any GPUs in the history had bigger pixel shaders than pixel pipelines, they used to be less or equal but never more.

[jAkUp]

Efficency is the name of the game, but going by those specs the G71 should lay the smack down once again.

One thing to consider, is the R580 does have the 512bit ringbus.

[Lfctony]

Yeah, because the Inquirer is very accurate.

"ATI R520 "Fudo" has 24 plus pipelines"
http://www.theinquirer.net/?article=22202

R520 has 32 pipelines, 24 working
http://www.theinquirer.net/?article=23648

ATI R520 scores 10K+ in 3Dmark05
http://www.theinquirer.net/?article=23752

ATI R520 has sixteen pipelines
http://www.theinquirer.net/?article=25730

Notice a trend there?

[nutball]

ATI are likely to have problems with 48 fat pixel shaders clogged in a thin 16 pipelines, never any GPUs in the history had bigger pixel shaders than pixel pipelines, they used to be less or equal but never more.

So what? Times are changing. Athlon I have to say that in all your posts here your views of what architectural features are relevant to performance seem to be stuck somewhere back in ... 2002 or so.

[Marvel_us]

Well Inquirer aren't the most reliable, true, but if I remember correctly they were pretty spot on with the G70 and 512mb G70 specs. Someone can correct me if I'm wrong.

Commenting on the card, it should be a hell of a nice upgrade from my GT.

Is there anyword if this will be using a 256bus or 512?

[jolle]

Well Inquirer aren't the most reliable, true, but if I remember correctly they were pretty spot on with the G70 and 512mb G70 specs. Someone can correct me if I'm wrong.

Sure they are right, eventually.. but they seem to often report every possible combination so that the sheer amount of their different theories sort of guarantees that they are atleast right once..

[Medion]

never any GPUs in the history had bigger pixel shaders than pixel pipelines,

Incorrect. The X1600 series does this. They use only 4 pixel pipelines and 12 shaders.

While the card is absolute ass at it's pricepoint, it compares favorably with many 8 and 12 pipeline cards. It's head and shoulders beyond any 4-pipeline card ever made. In other words, let's see how R580 and G71 compete when they're both out (if they both launch within the same rough timeframe).

[jAkUp]

Well Inquirer aren't the most reliable, true, but if I remember correctly they were pretty spot on with the G70 and 512mb G70 specs. Someone can correct me if I'm wrong.

Commenting on the card, it should be a hell of a nice upgrade from my GT.

Is there anyword if this will be using a 256bus or 512?

It have have 256bit memory bandwith, and a 512bit ringbus

[AthlonXP1800]

So what? Times are changing. Athlon I have to say that in all your posts here your views of what architectural features are relevant to performance seem to be stuck somewhere back in ... 2002 or so.

Look at Pentium 4's quad pump FSB, it not efficient.

Imagine ATI and Nvidia as water companies, ATI have 16 water pipelines with 512 bit ring bus, they had 48 days worth of pixel shader rainings in a month, it caused pipes blocky and it clogged the system and the end result is flooding, that not very efficent way. Nvidia have 32 water pipelines with 256 bit ring bus, they had 32 days worth of pixel shader rainings in a month, their 32 pipelines are withstand capable to cope with the water flowing.

[AthlonXP1800]

Incorrect. The X1600 series does this. They use only 4 pixel pipelines and 12 shaders.

Hmmm accorded to Beyond3D (http://www.beyond3d.com/misc/chipcomp/?view=chipdetails&id=104&orderby=release_date&order=Order&cname=) 3D chip tables, the X1600 (R530) have 4 pixel shaders.

[walterman]

Think that 48 PS with 16 pipes should be usefull for very complex programs, but, not all the pixels in the scene will use those complex programs, a lot of them will use more simple programs, so a wider but more simple design, will output more pixels in less clock cycles, in average. Also nvidia can borrow the extra full alu from the texture processor to perform 2 MADDs per clock cycle per PS, so it should perform like a 64 PS gpu. ATI can perform 1 MADD per clock cycle per PS.

I love 1900 MHz mem speed :)

[jolle]

Isnt staying at 16 pipelines going to hurt performance at very high resolutions?
What with all the "HD" fixation going on these days and all..

[coldpower27]

So right now current rumors dictate:

G71
32 Pipelines
32 Pixel Shaders
32 Texture Mapping Units
16 ROP's
10 Vertex Shaders? (would make sense)
256Bit Ext.Int. Memory Interface

750 MHz Core, 1.9GHz Memory

R580
16 Pipelines
48 Pixel Shaders
16 Texture Mapping Units
16 ROP's
10 Vertex Shaders?

695MHz Core, 1.55GHz Memory
256Bit Ext. Memory Interface, 512Bit Int. Ring Bus

Hmm there were some rumors before that R580 = 4xRV530

let take a look at that:

RV530 XT
4 Pipelines
12 Pixel Shaders
4 Texture Mapping Units
4 ROP's
5 Vertex Shaders
128Bit Ext. Memory Interface, 256Bit Int. Ring Bus

590MHz Core, 1.38GHz Memory.

R580 XTX
471% of RV530 XT for Pipelines
471% of RV530 XT for Pixel Shaders
471% of RV530 XT for Texture Mapping Units
471% of RV530 XT for ROP's
236% of RV530 XT for Vertex Shaders
225% of RV530 XT for Memory Bandwidth

Clockspeed is taken into account so keep that in mind.

I guess saying 4x RV530 is fairly accurate, since Pixel Shader performance is the primary performance determining metric in games, with the needed texture mapping units to back it up of course :),

[SH64]

Hmm G71's specs looks good & logical , but i'm leaning more towards ATI's approach by tripling up the shader power instead of increasing the pixel shader pipelines.
seeing how efficient the R520 with just 16 pipes/16 (MADD+ADD) shader ops againest the G70's 24 pixel pipes/48 (MADD) shader ops .. it scares me to think what will happen when you multiply that (R520's) by 3 & even increase the core speed esp in shader intensive games like FEAR.

edited : nevermind.

[walterman]

At 750MHz with 32 pipes, the G71 should be around 80% faster than GTX512 for shaders (750*32 / 550*24)

[coldpower27]

At 750MHz with 32 pipes, the G71 should be around 80% faster than GTX512 for shaders (750*32 / 550*24)

So this likely put's a single G71 "Ultra" at the performance level of the 7800 GTX 512 SLI.

[Graphicmaniac]

then it will cost 900$ lol

[aAv7]

Ok....1.) These are gonna be waaaay too pricey for me. $899 I gurantee...2.) If it doesnt have a 512 bus ouch. But they should. 3.) xbox 360 has been pwned.
4.) My f*cking 7800 gtx 256mb is considered acient..waaaaaaaaaaay too quickly.

When's the rumored release date for these? April?

[superklye]

So this likely put's a single G71 "Ultra" at the performance level of the 7800 GTX 512 SLI.
Just curious...what exactly are you contracting to get "put's"?

:p

[MUYA]

Hang on, are you all getting carrried away with teh pixel shaders thing? It's about what and how much of a particular shader op that particular pixel processing pipeline can do. Do not get carried away by the 48 pixel shaders etc or 32 pixel shaders, what they relate to is probably shader ops that can be done by the GPU. Tertsi has already givena reasonable guess here;

http://www.nvnews.net/vbulletin/showpost.php?p=773331&postcount=27

[walterman]

Well, here it's very simple, we took a G70 core, added a couple of extra quads, & raised the core speed. So, 80% faster. That's what the rumors dictate right now.

Of course, it's about the number of instructions that your chip can execute per clock cycle, like always. More ALUs, will let you run more operations without hazards on the fly. Like in any modern out of order cpu.

We're taking MADDs as a performance indicator, cause it's a common & expensive instruction for many shaders.

As Tersi said, each PS unit has several specialized ALUs, & a good shader compiler will spread the shader instructions properly among all the ALUs to keep all the transistors working all the time.

Perhaps nvidia could also tweak the ROPs for HDR+AA or other features, but, think that they said that want more general purpose power for the shaders.

[SH64]

Heres a rough comparsion made by a friend in other forum based on current rumors ... he assumed the G71 to be clocked @650/1800 (thought 750/1900 numbers are too much) , R580@695/1550 .

http://www.nvnews.net/vbulletin/attachment.php?attachmentid=15303&stc=1&d=1135964519

[ToxicTaZ]

Heres a rough comparsion made by a friend in other forum based on current rumors ... he assumed the G71 to be clocked @650/1800 (thought 750/1900 numbers are too much) , R580@695/1550 .

You should of did the chart with the RUMOR numbers! I have not herd once that the G71 would be clocked under 700MHz so the chart is not right!


Nvidia will clock the G71 higher MHz then R580.

[SH64]

You should of did the chart with the RUMOR numbers! I have not herd once that the G71 would be clocked under 700MHz so the chart is not right!


Nvidia will clock the G71 higher MHz then R580.

As i said its not mine , but anyway heres the numbers considering the G71@750/1900 (8 vertex units) :

Pixel fillrate = 12.00 Gpixel/s (G71 wins)
Texel fillrate = 24.00 Gtexel/s (G71 wins by large margin)
Multisample fillrate = 24 Gsample/s (G71 wins)
Geometery rate = 1.50 Gpolys/s (G71 wins)
Vertex shader GFLOPs = 60.00 GFLOPs (G71 wins)
Pixel shader GFLOPs = 384 GFLOPs (R580 wins)
Total shader GFLOPs = 444 GFLOPs (R580 wins)
Total FLOPs + texture address = 348 GFLOPs (R580 wins)
Main memory bandwidth = 60.80 GB/s (G71 wins)
Vertex texturing rate = 6.00 Gvertex/s (G71 wins as R580 assumed not to be able to do vertex texturing which is a standard SM3.0 feature. even if it did the G71 will still be faster thanks to the higher core clock)

i dont know how to calculate the total FLOPs with texture address .. i'll dig around.
anyone likes to modify the chart with the new numbers also is welcome! ;)

[walterman]

I dunno how he did the maths, but, i just disagree about the texture address gigaflops. Isn't the G71 going to use 32 texture units vs 16 in the R580 ?