AMD: Sea Islands R1100 (8*** series) Speculation/ Rumour Thread

yes, have seen it. But I tried without success to strike out "charlie" and insert Thomas Ryan. [-]..[/-] etc.. didn't work. So the edit took a while.

 

You should have tried it like that (it's the same on all vbulletin fora I visit):
[strike]Charlie D.[/strike] Thomas Ryan.

B2T, that stupid analyst should have asked, which 28nm HKMG processes they are going to use for what (GF's SHP [SOI] or HPP [bulk]?) and which one they are using for the current Tahitis (HP or HPL, my current uneducated guess would be actually HP, even if the voltages appears to be almost insane for that).

 

I tried to do it that way, but I use strike out nearly never, so... and google was no help.

 

So... approximately 15% more trannies? Well, I suppose part of that die increase could come from an extra 128-bit I/O to bring it up to 512-bit too.

AMD, Y U No Die Shots. .___.

---------

What's HSA?

 

HSA is the new legalproof name for fusion

 

Don't suppose we have any idea how many trannies a compute unit is, eh?

Hypa Shada Architek-cha!

 

HSA stands for Heterogenous System Architecture. I'll just quote the TechReport on what it stands for in a little more detail:

HSA won't mean much to discrete graphics chips initially though.

 

Thanks

 

I have a feeling we will be seeing a more efficient use of bandwidth rather than upping the MCs. It definitely is a possibility though.

So larger than R600(~420mm2) aka around 450mm2(+23% over Tahti) with a good increase in unit count or at R600 sized with a small increase in units but, possibly, even better clocks?

The nice thing about going forward with GCN is with the ROPs not being tied to the MC you don't need to try to increase clocks to makeup for staying with the same number of MCs, assuming you have enough bw.

 

What about increased density as well? By then they ought to know the ins and outs of the process so therefore ought to be able to stuff a few more in on the same space, right?

 

No. For that you need custom design and GPUs don't do that.

 

Well we did see much denser packing going from Rv670 to Rv770 presumably due to custom design for the compute units.

There's always the possibility that something similar could be done with GCN compute units assuming they aren't already custom. Highly unlikely yes, but potentially possible I suppose.

Regards,
SB

 

They weren't custom designed (not in the sense that people think of it anyway). ATI just unfuxxored their use of the tooling made available by the foundry, primarily. Such "free lunches" don't come all the time.

 

Ah, gotcha... Then yes, it's even more unlikely that GCN will feature denser transistor packing.

Regards,
SB

 

If Southern Islands / Radeon 7xxx series fails against Nvidia's Kepler lineup, how much pressure does that put on AMD to move forward sooner with Canary Islands (or whatever it's called) ?

 

dont see much how it can fail. looks like gk104 will be equal to 7970 at best. 7970 is smaller than 6970 so amd can drop the price down to 350 or lower (what 6970 was) no problem if necessary.

plus they'll get driver performance improvements and it's an overclocking monster. kepler could be an ocing moster too, but we dont know that yet.

even if nvidia has some "monolithic" gk 110 lurking, in a sense it isnt all that relevant, it will be equaled by the dual 7970 card, (which could also be equaled by a dual GK104 card of course, but not exceeded by a dual GK110 since that wont exist)

it may not be ideal but worse case for amd is no worse than last gen as i see it.