R600 has 320 shaders? HOLY ****!

[NV30>R300]

http://www.eetimes.com/news/latest/showArticle.jhtml?articleID=197700269

and

http://www.informationweek.com/news/showArticle.jhtml?articleID=197700288&subSection=Breaking+News


their is like many other articals as well comfiming this.

Was reading B3D just now and came across this...

I don't know what to say anymore.

[NV30>R300]

wow, I'm blind, mods move to roumor mill? lol.

[SH64]

Hmmm first we have to understand how these "320 multiply-accumulate units" work/compare to the 128 stream processors found on the G80 ..

[NV30>R300]

another source.

http://www.reghardware.co.uk/2007/02/28/amd_690g_launch/


Also B3D is leaning towards 320 scaler shaders.

[NV30>R300]

Hmmm first we have to understand how these "320 multiply-accumulate units" work/compare to the 128 stream processors found on the G80 ..


those "stream processors" are just simply a scalar shader with a missing mul.

If B3D is correcto, then R600's shaders will not be any different in a major way. That is both being scalar.

[Redeemed]

But in regards to shader power compared to that of the G80- going by the sheer number of "shaders" alone I'd think that R600 would stomp G80 in shading power. Would I be wrong to assume such?

[jAkUp]

Yipes. Well I will believe it when I see it.

Anyone know if the R600 is going to do anything with physics? Like NVIDIA's Quantum Physics? Will "Quantum Physics" games also be accelerated on an R600? (Havok games)

[NV30>R300]

But in regards to shader power compared to that of the G80- going by the sheer number of "shaders" alone I'd think that R600 would stomp G80 in shading power. Would I be wrong to assume such?


The number is all we know right now. If we do the math(I suck at it so bare with me), G80's shaders have a clock domain of 1.35ghz seperate of the rest of the core. If R600 is to not have a clock domain and the entire core is clocked at say... 850mhz, I believe R600 will dominate G80.

128 shaders @1.35ghz = 172,800 shader ops a cycle.

320 shaders @850mhz = 272,000 shaders ops a cylce.

That's nearly 100,000 more shader ops a cycle than G80. Again, that's assuming if they are the same in every way. They could be worse, or even better.

Now lets compare this to G81 that is said to have 25% more shaders.

160 shaders @1.5ghz = 240,000 shader ops a cycle.

still falls short of R600 by 32,000 shader ops a cycle. Now that assuming a 1.5ghz clock. Let's try 2.0ghz

160 shaders @2.ghz = 320,000 shader ops a cycle. A good jump over R600. Again though, this is assuming if Nvidia can hit that clock and if R600 does not have a clock domain. Now I'm only describing pure horse power here, not effiency. R600 could have a better lay out and arc that better feeds the shaders than G80. Then you have to take into account of R600 having thread packng.

[Redeemed]

I don't know, more and more this R600 seems to be shaping up much like the launch of G80- except the G80 was released on-time and available.

I'm hoping these new rumors are true, and that the R600 does stomp a mud-hole in the G80. I'd like that a lot. :D

[OWA]

Heh, I guess they had to put some number out there to keep people interested in the R600 and holding out for it b/c of the delay.

[tertsi]

uuuh noo hype hype.... the G71 has also "270 shaders" and it can do "175500 ops per sec" but still the G80 is like 2-4x times faster with same shaders... and btw. the R580 has "424 shaders". Peoples, do you even know what you're talking about? :)

Ok, Let's put it this way - the R600's *theoretical* (800 MHz clock, 64 4-wide MADD ALUs and no superscalar or co & dual-issue) MAD instruction performance is 51,2 GInstr/sec without any texture operations, while the G80 can do 50-51 Ginstr/sec with 8 texture sampling instructions.

Some results from my gpuinstr project...

G80 - MAD 3 512 8 256 100 132,4397 ms 50,6712 GInstr/sec (128 1-wide MADD ALUs or 32 4-wide MADD ALUs + some neat stuff - NDA)
G71 - MAD 3 512 8 256 100 241.2206 ms 27.8205 GInstr/sec (48 4-wide [vec3 + scalar] MADD ALUs)
R580 - MAD 3 512 8 256 100 289,5184 ms 23,1795 GInstr/sec (48 4-wide [vec3 + scalar] MADD ALUs)

[coldpower27]

We will see..:rolleyes:

[slaWter]

We'll see, in a year or so^^

[pakotlar]

uuuh noo hype hype.... the G71 has also "270 shaders" and it can do "175500 ops per sec" but still the G80 is like 2-4x times faster with same shaders... and btw. the R580 has "424 shaders". Peoples, do you even know what you're talking about? :)

Ok, Let's put it this way - the R600's *theoretical* (800 MHz clock, 64 4-wide MADD ALUs and no superscalar or co & dual-issue) MAD instruction performance is 51,2 GInstr/sec without any texture operations, while the G80 can do 50-51 Ginstr/sec with 8 texture sampling instructions.

Some results from my gpuinstr project...

G80 - MAD 3 512 8 256 100 132,4397 ms 50,6712 GInstr/sec (128 1-wide MADD ALUs or 32 4-wide MADD ALUs + some neat stuff - NDA)
G71 - MAD 3 512 8 256 100 241.2206 ms 27.8205 GInstr/sec (48 4-wide [vec3 + scalar] MADD ALUs)
R580 - MAD 3 512 8 256 100 289,5184 ms 23,1795 GInstr/sec (48 4-wide [vec3 + scalar] MADD ALUs)

nice math but ur wrong. it would be 64 4-wide + scalar. or 80 4-wide. if the r600 has clock domains, it should be able to do really good things. without them, it will merely be xx% faster than g80.

[J-Mag]

Anyone know if the R600 is going to do anything with physics? Like NVIDIA's Quantum Physics? Will "Quantum Physics" games also be accelerated on an R600? (Havok games)

Anyone Know if Nvidia is doing anything with physics? Any new news or just their original announcements? It is a paper feature...

[jAkUp]

Anyone Know if Nvidia is doing anything with physics? Any new news or just their original announcements? It is a paper feature...

I believe that Hellgate: London supports it.

[Zelda_fan]

I hope this is true. It's been a while since we've seen some competition from ATI/AMD

[Banko]

I don't know if this has been posted but it looks to be more like 160 because that test was run in Crossfire.

[J-Mag]

I believe that Hellgate: London supports it.

Damn I was hoping that it wouldn't be proprietary completely and that Havok would integrate with Quantum physics.

[NV30>R300]

I don't know if this has been posted but it looks to be more like 160 because that test was run in Crossfire.

how did you get this?

Separately, AMD gave one of the first public demos of the R600, its next-generation graphics controller that uses 320 multiply-accumulate units.

[tertsi]

nice math but ur wrong. it would be 64 4-wide + scalar. or 80 4-wide. if the r600 has clock domains, it should be able to do really good things. without them, it will merely be xx% faster than g80.

Like I said, *theoretical*...

64 4-wide + scalar (vec4 + scalar) MADD ALUs - 64 MAD4 instructions per clock - 51,2 GInstr/sec
80 4-wide (vec3 + scalar) MADD ALUs - 80 MAD4 instructions per clock - 64 GInstr/sec

Scalar doesn't mean nothing yet because we don't know can R600 really do parallel (co & dual-issued) different instructions... if it can, then....

64 4-wide + scalar (vec4 + scalar) MADD ALUs - 80 MAD4 instructions per clock - 64 GInstr/sec

Anyway... I think the big ?????'s in the R600 are texture sampling (address + filter) and special functions efficiencies. Hopefully R600's efficiency is better than R580 because the R580 takes roughly 25-30% performance hit with 8 TEX instructions :)

[SeriTonin]

But in regards to shader power compared to that of the G80- going by the sheer number of "shaders" alone I'd think that R600 would stomp G80 in shading power. Would I be wrong to assume such?

Don't say such things around here! ;) God forbid anyone give amd/ati any credit around here.

[retsam]

Don't say such things around here! ;) God forbid anyone give amd/ati any credit around here.
great, another brain dead idiot troll.

[El_Sid]

They need all shaders they can get:

http://biz.yahoo.com/ap/070226/amd_mover.html?.v=1

http://biz.yahoo.com/seekingalpha/070301/28448_id.html?.v=2

[retsam]

They need all shaders they can get:

http://biz.yahoo.com/ap/070226/amd_mover.html?.v=1

http://biz.yahoo.com/seekingalpha/070301/28448_id.html?.v=2
this is really a bad thing for the industry, i mean a really bad thing. would anybody want a bunch of pencil pushing bankers making technology desiccions instead of engineers? to me thats what would happen.