AMD: Volcanic Islands R1100/1200 (8***/9*** series) Speculation/ Rumour Thread

Retina iMacs are out in the wild - just run a benchmark or some kind of probing software (if there's such a thing) to verify what exactly is inside that GPU.

Anyway, "full tonga" hasn't ever been actually verified to exist, yes? No die shots that show additional shaders or such have been released, to my knowledge. So the tonga we have today could very well be the full one, and thus found in the new iMac.

 

HBM now available from SK Hynix:

http://i.imgur.com/8VjthWx.png

Direct link to PDF:

http://www.skhynix.com/inc/pdfDownload.jsp?path=/datasheet/Databook/Databook_Q4'2014_Graphics.pdf

Page (graphics memory pdf):

http://www.skhynix.com/products/support/databook.jsp

 

4Gb
128Mx32
8.0Gbps
H5GQ4H24AJR-R4C
FBGA(170ball)
16Bank, 1.55V/1.55V
Availaible Now

___________________________

1.55V, its a bit high or is it me ?

 

For the uppermost tier of GDDR5, this seems like the next incremental step. I'm curious if there's much headroom in that respect.

 

Claimed tests:





http://www.chiphell.com/thread-1182382-1-1.html

 

IF true, interessant, but lets wait and see..

 

Arrrr you inclined to believe this is real? Is the source (Junmiu) considered reliable? Or should we proclaim the official start of Silly Season?

 

Let's start with silly season and see where it goes, I have no idea whether Junmiu should be considered reliable or not

 

What are we looking at? Fiji graphs?

 

if the tests have any truth in them, the one who made the graphs seems to think it's the top of the line pirate islands based on the "captain jack" name, whether that's fiji or something else i have no clue

 

30% faster than 290X doesn't sound very credible for a device that's not in any way power constrained...

 

Yet that's more or less what GM204 is.

 

Well it is easy to do this type of graph or just modifiy one from a review.. Im allways really doubtfull when i see something without any photo or specs, nothing.

 

It would be awesome if true but I think a huge pinch of salt is needed here!

 

Speculation assuming true:

If it's true that's ~1.3x the performance using ~0.7x the power which is ~1.85x the perf/w if I'm doing this right - (65.6/50.1)/(197/279) - which is amazing. However, I would expect a bit more of a jump in performance assuming a new node increase surely? Then again the 7970 was about 1.3x the performance of the 6970 real world. Maybe they've just clocked it quite low to really try and decrease the power consumption?

 

Fiji is supposed to be manufactured on 28nm.

 

If that's the case and these results are representative then AMD have done amazing work.

 

I'm talking from the point of view that this is the massive new chip that AMD is supposedly working on, with 64 CUs and HBM.

You'd expect that one to be at least 45% faster due to CUs only, but actually faster because it has no memory BW constraints, yet here it's only 30% faster.

That could be explained by thermal issues, but that low power number puts that argument away.

Maybe it's just a cut down version...

 

A cut-down version is possible. I think the most likely hypothesis at this point is that this is just a fake.

But assuming it's not, AMD would have needed profound micro-architectural improvements to reach this level of power-efficiency. Thus, those CUs might not be directly comparable to, say, Hawaii's. Or they might be running at significantly lower clock speeds. After all, there's little point in pushing your chip's clock speed beyond its most power-efficient point if it's already faster than the competition by a commercially relevant margin.

 

Yes.

After Maxwell, it's clear that Kepler and GCN were not nearly as power efficient as we thought. AMD works within the same constraints as Nvidia. It's only normal that they were looking at ways to improve power efficiency as well, it's not as if they had another choice.

By the time this hits the market, gm204 is probably not where the puck will be.