Possibly.
It's easy for companies to plan when you have a steady cadence of process introductions. I have the impression that introducing 2 processes short after each other isn't doing the chip makers any favors.
AMD: Pirate Islands (R* 3** series) Speculation/Rumor Thread
- Thread starter iMacmatician
- Start date
-
- Tags
- amd
It's easy for companies to plan when you have a steady cadence of process introductions. I have the impression that introducing 2 processes short after each other isn't doing the chip makers any favors.
Then why didn't NV use 20SoC for anything GM107 up to GM206 so far? Considering A8X sports around 3b transistors, I wouldn't call it small as a chip either. In fact big enough to host up to mainstream GPU chips from either/or IHV in theory. A GM107 shrunk to 20SoC should be hypothetically in the 80-90mm2 league.
Either Apple has devoured too many capacities at TSMC for 20SoC or the rumors are true that 20SoC doesn't deliver as expected in power savings or simply both.
Since Qualcomm is also eating a big chunk of 20nm capacity, I think that's a very real possibility.
(Congrats on your new mini-Ailuros, by the way!)
With most of the cost increase coming from the fixed upfront costs(both 20SOC and 16FinFet), a minor increase in wafer price over 20SOC for a half-node like power/speed/density benefits of 16FinFet(+) is worth it.
Edit- More AMD sandbagging it seems like. 16FinFet was always expected in 2H '15.
16FinFet has a 15% improvement in density over 20SOC.
16FinFet offers 2x the gate density of 28HPM.
Somewhere around 10% seems realworld, maybe a bit more depending on ASIC design and process improvements.
Im keen to see what the next generation of AMD GPUs can do. If it comes out on 16nm finfet then we have a full process node + maybe the new memory architecture which should yield higher bandwidth. Maybe the reason why it is being delayed isn't the process, but instead a delay on the memory architecture?
Anyway my HD 7850 feels like it will need to be replaced by the end of next year, so I hope AMD has something for me.
Anyway my HD 7850 feels like it will need to be replaced by the end of next year, so I hope AMD has something for me.
While it was available for mass production, wafer availability was likely too limited for GM204 to be on 20nm. That aside, given the higher wafer costs and lower yields of 20nm at the moment, its cheaper to build it on 28nm until we hit the cross over point on per transistor cost (which is around Q1-Q2'15 as per Nvidia). Another point is the cost of masks is also significantly higher for 20/16nm.
Not all of AMD or Nvidia's chips are large (And the volumes are skewed towards the smaller chips anyway). And I wouldn't exactly call a ~130 mm2 A8X small either. Not to mention Apple's huge volumes. I dont have numbers on hand but I wouldn't be surprised if they actually use more wafers than Nvidia or AMD.
This I agree with. 16nm will be expensive and 20nm will be worth it for those who need the density increase from 28nm and don't particularly need the higher performance of 16nm. And 20nm wafers may actually be available in volume by the time 16nm is in mass production.
As I've mentioned above, aside from the limited wafer volumes, 20SoC does not bring any significant performance increase over 28HP, if at all. And given the higher cost per transistor, it made more sense to use 28nm.
16FF is more than a minor increase in wafer price. It is around 20% from what I've read. But yes the performance and power gains over 20nm are significant.
Yes 16FF was always on track for H2'15. TSMC claims mass production starting in Q4'14, compared to Q1'14 for 20 SoC.
From your first link, I don't believe he meant 15% density, his actual quote is
I believe the actual density increase is only ~5% as no-X stated. Thats the only credible figure I've seen.
Obviously it is good marketing speak but TSMC makes it sound like they can "quickly ramp" 16FinFet due to the similarity of the process. How quick they can switch over I guess depends on their current contracts and projected orders for 20SOC but if most of the current volume is the big players, they will want to switch to the bestest right away.
I haven't seen any actual numbers so I have to digress on the wafer price increase but the initial talk was if you can get the wafers you would be stupid not to go with 16FinFet.
The density increase of 16FinFet is like a ghost. I recall TSMC making one or two statements on it and giving actual numbers but I can't find the associated articles right now. They also tend to gloss over which exact 16FinFet they are talking about, I've seen statements made about 16FinFet increasing density over 20SOC and 16FinFet+ increasing density over 16FinFet.
It all seems a bit murky and I think will be highly dependant on how well the process is implemented into the specific designs.
Yep..thats what TSMC has been saying all along..and they even claim that 16FF is ahead of schedule (by just a quarter though). As the backend is similar, they should be able to ramp it quick enough, though yields may trail 20nm for a while.
I'd posted some figures in the Maxwell thread a while back..I'll reproduce those below:-
- Cost of a 28nm wafer - $4,500-$5,000
- Cost of a 20nm wafer - $6,000
- Cost of a 16/14nm Finfet wafer - $7,270
Yea it does seem a bit murky and there have been some conflicting statements made (Possibly due to misinterpretation) but I recall two figures direct from TSMC. They claim 1.9x density improvement for 20SoC over 28HPM and 2.0x for 20SoC over 28 HPM, i.e. ~5% for 16FF over 20SoC
For FF+, the only concrete statements I can find point to 15% increase in speed over FF, nothing about density. However, I would not be surprised if there is a slight increase thanks to process refinements.
hm... Mobile Radeon refresh
(and consoles)
It could be possible that AMD has opted for 20nm..but given the path NV has chosen, I wouldn't be surprised if they're going straight for 16FF either. I believe that's the strategy for their APU's as well.
On the console front, yes..20nm should reduce costs and power..they dont need more performance. But I dont see it happening until the cost per transistor and wafer availability are much better. Earliest I can see it happening is late 2015.
Some time ago, there was this leaked slide from a CAD vendor listing 2 GPU tape-outs for AMD. One was >300mm2 (Tonga most likely) which was listed as 28HPM. The other one was >500mm2. Does anyone remember if that was 28HPM as well? If so, the whole waiting for the process to be ready definitely doesn't fly.
Edit: here it is: http://www.kitguru.net/components/g...-gpus-made-using-28nm-hpm-process-technology/
Edit: here it is: http://www.kitguru.net/components/g...-gpus-made-using-28nm-hpm-process-technology/
By the way, this tape-out apparently happened in H1'2014; it would be rather odd for the actual product to be delayed into H2'2015.
Fair enough. Thanks for all that.
Saving it for Spring '15? Was wondering if they were spacing products out if they weren't confident about the timeline for the next node.
DuckThor Evil
Legend
We already saw this:
and if the chip taped out long time ago, it'd be weird if it only ships H2 2015... Either it's coming out sooner or it's not coming out on 28nm imo like it was supposed to. Launching something like that against Maxwell Titan might not look too good. That 500mm2 AMD chip seems like it has no chance against the big Maxwell in performance and would consume a ton more power. I mean hypothetically if it's 25% slower and needs a liquid cooler, while the big Maxwell runs cool on air, it looks bad.
and if the chip taped out long time ago, it'd be weird if it only ships H2 2015... Either it's coming out sooner or it's not coming out on 28nm imo like it was supposed to. Launching something like that against Maxwell Titan might not look too good. That 500mm2 AMD chip seems like it has no chance against the big Maxwell in performance and would consume a ton more power. I mean hypothetically if it's 25% slower and needs a liquid cooler, while the big Maxwell runs cool on air, it looks bad.
That's really not how semiconductor companies work: when you have working silicon, you ship it. There's never a quick fix, but there's always somebody who's not going to care too much about power consumption as long as performance is there. There were plenty of people who bought GTX480 or even R600.
If it gets a decent cooler, like the water cooler that was posted, it should have an audience that's willing to buy.
Tonga may be a disappointment to many, but it's not a bad chip and IMO doesn't deserve being use as an example of power savings not working out. It's just that it's no match for the competition...
It also isn't much of an improvement upon Tahiti, at least in the R9 285. Sadly, it's already been two months and there's still no sign of the 285X, so that's the only data point we have.
For the Radeon Gaming product line...
Well, I haven't seen any benchmarks of Tonga outside of it. I might have missed them.