New AMD GPU Family Codenames "Volcanic Islands" and "Pirate Islands"

[HumanSmoke]

I just hope they can manage atleast 30% more performance than Titan with the new manufacturing node, if it's less then Nvidia will probably charge over the moon for Maxwell...

If its 30% more than Titan then we're fucked. AMD will use the same pricing strategy as the initial HD 7970 launch ($549) and the HD 7990 ($1k) and price according to the competition...then we have to rely upon Nvidia lowering prices or AMD suddenly deciding to not cash in on...a cash cow.

30% more performance than Titan means roughly 50% increase over the 7970. Even with a tweaked architecture, that's going to be a big-ass die considering the process node is still 28nm and heat/power are still going to limit core clock to what is presently in use.

 

[radrok]

You sure that's going to be on 28nm? If so, well I don't think they can manage much more than just matching the Titan unless they shoot well over 300W TDP.

 

[HumanSmoke]

You sure that's going to be on 28nm?

Well, TSMC themselves have said that (commercial) production won't take place until the new year, soooooo if AMD are doing the big reveal in September either its going to be the biggest tease/longest paper launch in recorded history, or its a 28nm GPU.

If it is 28nm which seems the more likely, then your ~300W estimate may prove to be on the money. Can't see AMD stripping out features like double precision to save die space. I also can't see AMD risking a "Fermi" situation with a huge die on a largely untried 20nm process unless yields are spectacular- and if yields are that good I'm pretty sure TSMC might be already yelling that from the rooftops.

 

[TheoneandonlyMrK]

switching to different process (gate last vs gate first) would cause a hell of a headache and definitely not cheap. i think amd would keep glofo for cpus only. it's a bad (and done) deal anyway.

Seams process variation is an end goal in modular ip construction though why make just drop in tools and not node span them ie the job might not have been as big or may have started way sooner than we all think Si was gcn2 and it disappeared Vi is alleged to be a new architecture and I've not seen anything detailing Vi at all in terms of an arch reveal could Si have been binned the minute the gk104 turned up.
I got some special day dreaming skills eh but anything almost is possible.

 

[Casecutter]

Kind of depends on what you're expecting

Well I meant it's not as problematic departure as would be from VLIW to GCN all on a die-shrink. There's hopefully less driver instructions to wring-out and what not.

the expectations of others (~25% increase in die size, core/TMU/ROP count, TDP etc. seems to a favourite configuration),

I'm not subscribing to that, I don't think they need or want that. A 25% bump makes a die that ends up only 13-15% smaller that a GK110 or perhaps about what's "used" in a GTX 780. I see that as really starting to hurts piece price. I don't see that as necessary, while I think raising TDP is not something AMD wants to elevate. All AMD needs is 18% to be at a GTX 780, or 27% for Titan, while hold TDP and per chip cost down.

Although, knowing what Bonaire (GCN 1.1) provided on a 30% die increase, with about 30% in transistors, 40% increase Sp and TMU’s, though no bump in ROP's. That provided a 20-25% in performance, while a 20% lower watt/perf from the 7770. However, with artificially low clocks and 128-Bit I think there was a bunch of under-utilized potential leftover.

Now I could be wrong, but honestly I sense Tahiti was not all that optimize for the size (ie. more ROP’s?). I'm holding out hope there's a still a good amount of juggling of that GCN to pull more performance (25%) without any major bump in die size, TDP, and perhaps three true product derivatives. I don't believe AMD need/wants to build a chip that escalates what Titan does, "just best it" at a price that makes bringing a "full GK110" not a competitively priced alternative. But then knowing Nvidia they'd release it just to claim the "Crown" no matter the price.

 

[HumanSmoke]

Now I could be wrong, but honestly I sense Tahiti was not all that optimize for the size (ie. more ROP’s?). I'm holding out hope there's a still a good amount of juggling of that GCN to pull more performance (25%) without any major bump in die size, TDP, and perhaps three true product derivatives.

Soooooo, you are under the impression that AMD can deliver a 25% bump in performance over Tahiti just by tweaking the uncore, without altering power usage, frequency, and die size. This should be fun to watch play out.

 

[Casecutter]

Soooooo, you are under the impression that AMD can deliver a 25% bump in performance over Tahiti just by tweaking the uncore, without altering power usage, frequency, and die size. This should be fun to watch play out.

I'm holding out hope there's a still a good amount of juggling of that GCN to pull more performance (25%) without any major bump in die size, TDP, and perhaps three true product derivatives.

Well without any major bumps (<5-10%) I never said frequencies. If all they do is bump the die and power 25% to get 30%, that's a no brainer.

3870->4870 had even held to the same 55nm process, but it really was like 36% increase. But that I believe was more what GDDR5 brought

We’ll see… without a die-shrink I might well be overly-optimistic. As the RV670>RV770 had a 30% increase in die-size and used 25% more power, while a the TDP jumped by 40% even having DDR5... Ouch! Although, I still think there might have been "aha! moment" with GCN and perhaps why AMD seemed to hold-off not releasing 8XXX series stuff. Given the price we see some Tahiti cards working down to in price, and perchance better yields AMD can bump the die size and still offer good price.

The whole speculating and/or reading the tea leaves is just recreation. We watch and B.S. what we see/feel.

 

[radrok]

Why am I under the impression that their flagship is going to be 250W+ close to 300W?

If it's still 28nm I'm gonna bet those power figures are plausible.

 

[Casecutter]

Looking a W1zzard on Titan Peak power numbers (while gaming) the difference between the original 7970 and Titan was 25%, while performance/Watt where equal. That would indicate that that if Titan falls in as a 250W TDP, I would think it's not that far of stretch to have this hold to something closer to a 250W TDP also.