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

"Too much" as in "too much money being spent on memory for a $250 graphics card".
Specially when the top-end Fiji - which is more or less 2 Tongas glued together + HBM - only gets 4GB.
 
Actually, G92's longevity has already been beaten by Pitcairn with a sizable margin.
 
Actually, G92's longevity has already been beaten by Pitcairn with a sizable margin.
Is it though? Has the matter of "why doesn't Pitcairn support OpenCL 2.0 while this-supposed-Pitcairn does" been solved for example? Having same "core configuration" doesn't make it the same chip
 
I've seen multiple confirmations that it is the same chip (dimensions, firmware ID, exchangeable firmwares between older and newer models, etc.).
As for OpenCL 2.0, I see no reason why the 2012's Pitcairn couldn't support it through hardware, so it's probably just driver support attached to model ID.

Besides if you go that way, G92 had a process shrink to G92b, which should be considered a much larger change than whatever they could've done to Pitcairn IMO.
 
Yeah, that's what I mean by opening the door for X SKU. :)

No, the 384bit bus is disabled because it requires a crossbar for the ROP/bus mismatch and will not be enabled. Nor will any old card be rebranded next year, at least nothing $150+. The 380x is just a buildup of 380 cards they didn't need to bin down because all of the silicon worked. It'll come and go briefly, being available in whatever quantities are made incidentally available for however long the GPU run lasts.

For perspective, at 14/16nm and AMD's goal of doubling performance per watt a gpu equivalent to the 380x should, as a brief estimate, have an MSRP of $179 or less. The next node is a huuuge jump for both AMD and Nvidia, and any thoughts of rebranding old cards on older nodes somehow competing with GPUs on the new node is ridiculous.
 
Need say, that the last years was a bit complicated by the fact that 28nm have last very long.. really more time that it was previously expected .. both AMD and Nvidia have ned to revisit their expectations with more or less success with it.

2016 should be a complete turnaround on the respect of what we have seen thoses last 4 years.. I expect the jump to be really high, and from what i have seen on both side, this look to be the case, but maybe it is, this time, me, who put my expectations a bit on the high side ..
 
Last edited:
Radeon R9 380X launched.
Review links shamelessly copied from wccftech:

At the moment, it seems to bring more value than the GTX960 4GB for the same IMO. At least until we put the GM206's giant overclocks into the equation, that is (I imagine the GTX 960 overclocked towards 1.5GHz is bound to give the 380X a very hard time).
But at least the Asus card seems like a little bit of a missed opportunity. The cooler seems great as it keeps the card at a maximum of 69ºC while being very silent, and the power consumption is lower than last year's R9 285. It seems to me that they could up the core/memory clocks some ~10% by default and allow the temperature to rise up to ~80ºC, at the cost of higher load power consumption. Anyone who's terribly concerned about power consumption will go for a GTX 960 anyways.
 
Pretty meh card with price midway between 380 and 970/390 while performance pretty close to 380, sometimes only around a 5% improvement.

Lower clocks have allowed them to cut down on power however, with strix 380X taking 50W less for a bit better performance than 280X in TPU review. 960 also looks soundly beaten, with the vanilla 380X being 28% faster at 1080p and also overclocking 17% on core basically nullifying the OC advantage that you find in other X card scenarios with AMD clocking them sky high.
 
That's still not the full Tonga, right? If I remember correctly, the presence of a 384-bit bus was confirmed some time ago. This is pretty disappointing.
 
That's still not the full Tonga, right? If I remember correctly, the presence of a 384-bit bus was confirmed some time ago. This is pretty disappointing.
That's correct. They've only just enabled all CUs.

So I guess the question for me is what impact there is to overclocking the memory.
 
I suspect that AMD did not reintroduce a ROP-MC crossbar, so that means we only get 32 ROPs out of the 48 that are likely present. That kind of sucks. I don't understand why AMD would do something like this, they're hardly in a position to sandbag.
 
I was leaning towards 48ROPs too, but if it made an appreciable performance improvement they would have done so by now.

380X's power saving itself might be due to Asus version, their Fury version was lower clocked than Sapphire's and consumed less too. With the price tag on it, this card should've been out during Fury and 390 release.
 
I suspect that AMD did not reintroduce a ROP-MC crossbar, so that means we only get 32 ROPs out of the 48 that are likely present. That kind of sucks. I don't understand why AMD would do something like this, they're hardly in a position to sandbag.

Maybe because a 48 ROP + 384bit + 6GB GDDR5 + color compression combination would get Tonga uncomfortably close to the R9 390 8GB, especially if given enough voltage to reach 1100MHz at standard speeds.
Though we may still see such a card, if/when AMD runs out of Hawaii chips that need binning.
 
The crossbar wouldn't be needed if there were 48 ROPs. It existed in Tahiti because of the extra memory channels.
The most recent GCN revisions have been described as capping the number of render back ends per shader engine to 16, and it doesn't look like the shot of Tonga has a third engine.
Is there an extra 8 blocks for ROPs in the shot? I think those might be the row of paired rectangles (depth and color?) that line up alongside the CUs to the left and right of the CU array.
 
Back
Top