AMD ditching VLIW and going SIMD with Southern Island.

[notty22]

Well more than a couple sites are running with it, as the next gpu design.
I guess its up to AMD to correct perception ?
AMD announces new architecture. Release date:[URL="http://www.google.com/url?sa=t&source=web&cd=4&ved=0CC4QFjAD&url=http%3A%2F%2Fwww.gamespot.com%2Fforums%2Ftopic%2F28719165%2Frumor-amd-announces-7xxx-series.-release-date-next-month&rct=j&q=amd%20announces%2079xx%20gpu&ei=xzX-TcODAabi0QGfjvHEAw&usg=AFQjCNGSzKCzf92g3UY4-gm9DD9Ml7i12g&sig2=98JIPQSvziIlVGp_uCGJLw&cad=rja"]Later this year ...
[/URL]



AMD Radeon HD 79xx Series GPU Details

 

[Dribble]

As a programmer who's used both CUDA and AMD OpenCL, I'm very pleased with this development. Whenever you program for a GPU, you have to envision your code running on dozens or hundreds of processors in parallel. But to get maximum performance out of VLIW, you had to create vectors of four integers within that already parallel architecture! SIMD should eliminate this additional complication.

For those worried about AMD performance lowering, don't be. The biggest current nVIDIA GPU has 2x512 processors. AMD has some GPUs with thousands, albeit at half the clock speed. This should eventually make AMD more competitive for things like BitCoin...or PrimeGrid.

It depends how you measure it, a more accurate measurement for comparison with nvidia numbers would be to divide it by the VLIW number - so for example the 5870 actually has 1600/5 = 320.

As for bitcoin speed as I understand it radeons are good because of huge integer performance on simple operations due to the fact they are basically designed for DX9 graphics, which also just happens to be what bitcoin mining requires. If they want to compete with nvidia on more complex compute operations they'll have to change that which means less die area for that type of integer math, which means comparatively lower bitcoin (and probably DX9 graphics) performance.

As for the release dates people are bouncing around - well it's obviously going to be a huge chip on a brand new node size. See fermi for how well that worked out. Surely it's more likely they'll go for a simpler, better understood VLIW4 design at 28nm first, then move on to this new one when things have stabilised.

 

[lifeblood]

There has been no official source that confirmed HD7000 to be a GCN GPU. Given that AMD is switching to a brand new untested 28nm tech, do you think they'll also throw in a brand new architecture on it as well? Seems like too much risk to me. For this round, they could just increase the current specs of 6970 and be competitive. But who knows.

Agreed. nVidia got burned big doing this. I can see the next version (HD8000) using the new architecture on 28nm, but not the HD7000.

 

[Kenmitch]

Agreed. nVidia got burned big doing this. I can see the next version (HD8000) using the new architecture on 28nm, but not the HD7000.

Although it could go either way I'm leaning towards a combo of both new/newer from AMD once 28nm is fired up. For all we know at this time AMD has working silicone already. Might be fire breathing at 40nm tho. It's possible that this new design will be geared towards a new market for AMD and not a catch all at first. On another note it's also possible they could whoo intel on 22nm for production. Time will tell.

 

[Dark Shroud]

On another note it's also possible they could whoo intel on 22nm for production. Time will tell.

I've been thinking the same thing, Intel opening their 22nm fab up for AMD's GPUs just to take Nvidia down a peg.

 

[Kenmitch]

I've been thinking the same thing, Intel opening their 22nm fab up for AMD's GPUs just to take Nvidia down a peg.

I didn't imply anything other than maybe just the fab tech for maybe some help on Intels gpu's as we all know intel needs it. But now that you brought it up it could be a double win for intel....That's only if intel has any interest in owning nvidia tho.

 

[Maximilian]

HD5xxx = SIMD ?
HD6xxx = VLIW ?

Is this correct?

 

[Cerb]

HD5xxx = SIMD ?
HD6xxx = VLIW ?

Is this correct?

No. AFAIK, Radeon 9500 Pro (R300) to 6xxx are SIMD VLIW.

 

[Lonbjerg]

No. AFAIK, Radeon 9500 Pro (R300) to 6xxx are SIMD VLIW.

R300 to to Juniper is 68xx is VLIW5
69xx is VLIW4.
http://www.anandtech.com/show/4061/amds-radeon-hd-6970-radeon-hd-6950/4


Fermi is SIMD

"SIMD VLIW" is so wrong to say that I don't know where to begin.

 

[ViRGE]

R300 to to Juniper is 68xx is VLIW5
69xx is VLIW4.
http://www.anandtech.com/show/4061/amds-radeon-hd-6970-radeon-hd-6950/4


Fermi is SIMD

"SIMD VLIW" is so wrong to say that I don't know where to begin.

Technically AMD is making a SIMD of VLIW units. It's not a traditional vector SIMD in any sense of the word though, which is why GCN is such a different design.

 

[Cerb]

"SIMD VLIW" is so wrong to say that I don't know where to begin.

Would adding another term to the mix, SIWMD, make it better? D:

 

[Dark Shroud]

Stop it before someone in marketing reads that by accident.