AMD: Pirate Islands (R* 3** series) Speculation/Rumor Thread

Ah I didn't realize the alliance had diverged to such a degree. Thanks for the clarification.

 

From a cost/transistor perspective, it makes more sense to have the lower end parts on 28nm. Not to mention availability and yields.

However, I could see one area where they are eager to move to 20nm fast. ARM Server. They claim that their 28nm part is on track for Q4. Cost/transistor is not such an issue in this segment and power is more important so I'm sure they'd wanna move to 20nm as soon as possible.

Haven't heard any news on GM200's tape out yet.

I hope I'm at least partially correct as well And yea I was thinking that as well..those comparisons would be extremely useful and interesting to see.

Yep the only difference is that Samsung claims that its process has a higher density than TSMC. TSMC's 16nm process has a very small density improvement over 20nm, ~5%. Samsung claims a 15% improvement for their 14nm process.

 

Did Samsung directly reference the density of TSMC's 20nm?
What if TSMC's 20nm is denser than Samsung's? Lesser scaling from a better starting point can still lead to the same end result.

Granted, as noted before, there is a metric ton of PR garbage inherent to any density discussion.
The headline density numbers may vary depending on what structures are being used as the metric. Less regular logic, a global average, or highly regular SRAM would have different values. Do we know which PR statement is the source for both?

Headline density may or may not be used in a given product, and marketing pressures can lead to the quoting of higher density numbers that some products would avoid using.

The other question is whether TSCM's or Samsung's 20nm are 20nm, or the same value that is not 20nm, or even if they're talking about the same critical measures, or if those measures are not offset by other unmentioned quirks or side effects.

Or if the 16/14nm metrics are even consistent within manufacturers, much less between them.

 

I wonder why there's so much marketing bullshit when foundries' customers are experts who ought to be immune to it.

 

Probably the managers are still the ones ultimately deciding on which foundries to fab products in. If the number of experts in their company is diminishing or if they are a horde of proverbial engineers which are not highly skilled in communicating..

So maybe the foundries are on to something

 

Perhaps not every team is wholly immune, or maybe the execs above the engineers aren't.
Their stockholders, and their customers' stockholders definitely aren't.
What else are can they say?

At any rate, as briefly interesting as a stagnant node generational comparison might get, I'd rather the designers nut up and do something interesting instead of another .1 increment to the "just finally not embarrassing" stuff they released years ago, or freaking Llano, blech.
Node transitions aren't a guarantee of courage, but at least there might be something akin to movement that isn't an auto-increment script to the numerals on the box or another voucher.
At least the scrypt mining bubble had the bonus of being a farce.

That's not to say that it's a good enough economic argument. What bores or excites me isn't going to drive them much in the way of revenue.

 

When Morris Chang rebutted Intels claims about TSMC scaling, he used the following graph. link

As has been said already, the density improvement depends on your assumed mix of features, so grains of salt are recommended. Nevertheless, this is TSMC themselves showing the 16nmFinFet process being a small step in terms of density (but still a step), whereas the 16nmFF -> 10nmFF is a full nodal change.

On a grander scale, plotting process advances vs time over the last 20 years paints a dire picture, and it won't get easier the next decade. Lithographic advances will diminish in importance as a driver of performance as we progress in time. Remains to be seen where that leaves the industry. The IC business as a whole is still growing at a healthy clip, scaling woes notwithstanding.

 

Is TSMC comparing Intel 14nmFF to their 16nmFF or their 16nmFF+?

 

Since we now know that AMD will not have any GPUs on 20nm this year can we concur that all of the other data from eXtremeSpec is just made up?

 

28 nm is not of interest I guess to anyone anymore but if those are on 20 nm (let's not write them off just yet) and with a little bit more imagination to the speculation, we can guess that those are tape-out dates or kind of.

 

Are we going to see stacked DRAM from NV and AMD this year?

 

Definitely not this year, at earliest, late 2015 from everything I've read.

 

AMD to Launch New Single-GPU Card This Summer, to Take on GTX 780 Ti
http://www.techpowerup.com/200721/a...u-card-this-summer-to-take-on-gtx-780-ti.html

This news post from a few days ago is what got me thinking about it.

 

That page gets things wrong about HBM. It states it will have 128 bits per stack, which limits things to 4 stacks on a card because the GPU has a 512-bit bus.

A single layer in an HBM stack has 2 128-bit channels, with a 4-high stack having 8 channels/1024 bits.
A single stack by itself is more than what is claimed to be possible, and until a higher density revision of the DRAM is available, 4 stacks for an aggregate 4096-bit memory bus would be necessary to keep capacity the same.

 

I thought the 290X was already doing a fine job taking on the 780 Ti? If they do manage to get something out appreciably faster in the short term than either of the former cards that would be impressive.

 

Makes it sound like the r290x wouldn't already compete with the 780 Ti.... Last time I checked it was within 5% of the performance of it, unless you use the reference cooler on the quiet setting.
In any case seems doubtful to me we'd see HBM with chips already on the market.

 

I think this rumor is much more interesting: http://videocardz.com/50506/exclusive-amd-working-tonga-gpu

 

Check CodeXL 1.4 PDF
you will see Volcanic Island vs Hawai Sea Island (ASIC)

Check C++ AMP Lambda API, which also for Xbox One
You will see AMD Next Generation Tiled GPU vs older Untiled GPU

AMD will moved to Tiled Architecture

 

Its been what 6 or 8 months since the r290x was released ? Better yields plus a water cooler would most likely make for a very fast single chip design from amd

 

Only reference to Volcanic Island, which Hawaii should be too, is that 1.4 adds support for Volcanic Islands.

Where exactly?