Raja Koduri of AMD confirms two new 14NM/16NM GPUs next year

[Boomstick777]

not sure they will reuse the Fiji core as it is too big, maybe re-spin it on the smaller process but then it's a new core anyway, really.

Yeah expecting a re spin, but reuse existing Fiji / HBM1 for those parts. So not a true new card if you get my meaning. Whereas the two they discuss in article could be entirely new with HBM 2.0 with the rest of the stack being re spun and getting a bump.

Fury X could occupy a lower down slot, below the 2 top of the line new cards.

This could make sense imho. Next year is going to be great for gaming man, not so great for my wallet..

 

[Orangey]

One theory I saw bandied about is a bunch of small dies stitched together on an interposer.

 

[Marine-RX179]

So, two new cards and the stack gets a bump up, i.e some rebrands to occupy the new mid - high range, Fury X becomes new mid - high range. Top high end is new part based on 14nm/16nm..

Fury X replaced by >> 14nm/16nm GPU 'Full fat'
Fury replaced by >>14nm/16nm GPU 'Cut down'
390X replaced with >> Rebrandeon Fury X
390 replaced with >> Rebrandeon Fury
Nano >> Rebrandeon Nano..

Even if the Fury cards get rebranded, are we sure they will be rebranded in the form of 28nm, rather than 14nm/16nm?

I mean on 28nm it'd probably cost AMD more to make the GPU...if they can move the Fury design onto the smaller die, it would cost less to make, and allow room for improved efficiency and temp or increase the clock speed...similar to how 9800GTX 65nm to 9800GTX+ 55nm back then with the die shrink?

 

[nashathedog]

So two new cores, to quote the original Forbes article.



We have to assume that it will be a new top end core with two or three cards coming form it, and the next tier down with another two or three cards from that one, that would leave maybe Tonga the currant 380 or the new renamed Antigua core 380x to fill in the lower end parts. It will probably mean another run out for Oland or Bonaire for the ultras low end, Overall it should work out ok.

If the performance hike a lot of people here claim is coming is true won't that be Grenada and Fiji to fill the low to mid end with Tonga at the bottom of the pile?

Even if the Fury cards get rebranded, are we sure they will be rebranded in the form of 28nm, rather than 14nm/16nm?

I mean on 28nm it'd probably cost AMD more to make the GPU...if they can move the Fury design onto the smaller die, it would cost less to make, and allow room for improved efficiency and temp or increase the clock speed...similar to how 9800GTX 65nm to 9800GTX+ 55nm back then with the die shrink?

That makes more sense and is probably what we will get when they say new 14/16 cores. (Mid level)

They'll do as Nvidia did with the 980/970, release a shrunk Fiji first as it will be faster than the current models and then later on release the new high end cards.

 

[flopper]

If the performance hike a lot of people here claim is coming is true won't that be Grenada and Fiji to fill the low to mid end with Tonga at the bottom of the pile?



That makes more sense and is probably what we will get when they say new 14/16 cores. (Mid level)

They'll do as Nvidia did with the 980/970, release a shrunk Fiji first as it will be faster than the current models and then later on release the new high end cards.

600mm die today 28nm will be around a 300mm die with the new 14/16nm.
a 400mm die using 14/16nm tells us if they are able to do that we will see a huge leap due to the die shrink alone which historically is the main reason for progress.

 

[Dygaza]

Makes no sense rebranding Fiji chips, they're far too expensive for mid tier.

 

[Mei]

its still mabe 6months away tho?
bit soon for getting excited
unless they want to send me one early then ill get excited

 

[LoadsaMoney]

Price is irrelevant, if it performs it sells

Pricre is relevant when its AMD.

 

[Geeman1979]

Pricre is relevant when its AMD.

Agreed ^

 

[Seanspeed]

Fiji is a huge core and has the most to be gained from in die size/cost over any other core. Not least because interposer/HBM will be the strongest growth area in terms of design knowledge. It's the first real mass produced interposer based design so there will be a lot of information on how to make it better, get better yields, but also as said die size means it's expensive.

99% likely the two cores will be in the 200/400mm^2 areas, with a future Fiji replacement to come sometime in 2017.

So 7970/7770 type cores all coming say July/August, then a big core coming after yields improve, volume ramps and prices per wafer and per chip come down as a result.

Low end, low end is so low power, cheap, small margin and increasingly small volume that there is incredibly little to be gained by making a new small core on a highly expensive process. So some 70-120mm^2 sized low end thing isn't a priority, useful eventually as a discrete gpu with a bit more power in some laptops... if apus don't start coming with HBM in which case low end discrete gpu for laptops disappears as entirely pointless.


I expect the same from Nvidia, GTX 680/660 type cores in 2016, a Titan level core in 2017.

Cool, thanks for the insight.

 

[Boomstick777]

Makes no sense rebranding Fiji chips, they're far too expensive for mid tier.

In that scenario, Fiji would be re spun on a die shrink. Costs reduce over time and with tech advance. Fiji could become new mid / high tier and the two new cards would occupy the new top spots. This would be in line with previous generations, it would make sense.

 

[flopper]

the normal pattern are to produce a small die on a new node.
however todays apu etc..are so powerful low end is dying out.
its more likely if yields are good to go with a bigger die to create some momentum.

 

[Mtom]

Aren't there Greenland, Baffin and Ellesmere rumored?

 

[bru]

Aren't there Greenland, Baffin and Ellesmere rumored?

One of those might be the dual GPU variant, so still only two new cores.

 

[LoadsaMoney]

If Fijis reused, it'll still be HBM1 won't it ?

 

[Orangey]

I think the signalling is the same so they might be able to drop gen2 chips in place.

 

[drunkenmaster]

Fiji will categorically not be respun and makes zero sense being continued on 28nm when even a 300mm^2 part would likely beat it by a good 10-15% on 16nm due to new architecture/efficiency.

You have to tune bits of architecture to a specific process. The new architecture has to be done on the new core, doing the old architecture on the new process is another very large piece of work and cost. It would be, because there is already a new architecture taped out on the process, far cheaper to make a circa 300mm^2 core with the new architecture than to shrink Fiji which would required redoing the old architecture on the new process for a single chip.

New architecture will incorporate everything they learned on memory controller and architecture in general optimised for HBM as well as multiple other improvements and newer features. Fiji redone on 16nm would be effectively the worst decisions AMD could make.

Even Nvidia after years of problems stopped attempting 500mm^2 + cores on a new process. GK100 both came later on and even when it arrived was very low yield and put into higher yield/lower volume professional lines. I'd be surprised to see a big core from either company. This leads you to think anything from a disappointing 150-200mm^2 for one of these gpus and 300-350mm^2 for the other. While it should comfortably beat Fiji that wouldn't be awesome in comparison to Fiji/980ti. A little more optimistic and probably more likely is maybe 200-250mm^2 for one gpu and 375-450mm^2 on the other. a 450mm^2 die could be with architectural improvements, potentially a good 60% faster than Fiji.

I would honestly put 500mm^2+ cores at an extremely low chance and may not even appear later on either. 10nm is a bit of an odd one. The shrink isn't really what slowed everyone down directly, it was double patterning and finfets more than the shrink which is why 10nm looks to be pretty close behind 14/16nm, because once they crack the finfet/double patterning, it's really the same tech being used for 10nm.

So rather than circa 400mm^2 core then a 500mm^2 core 9-12 months later, we might see 400mm^2 16nm, then 300-400mm^2 10nm 12-15 months later instead.