AMD: Speculation, Rumors, and Discussion (Archive)

[Deleted member 13524]

There are some pretty compact GDDR5-based boards, and Polaris would be sporting a 256-bit bus like the 970 mini.

That's right. And although the 970 may have a somewhat simplified memory footprint due to the 224bit bus, I doubt that taking away 1/8th of the connections made a huge difference.
Then again, we never saw a mini 980.. There's a "980 for laptops" but I don't know how big it is.

 

[CarstenS]

There is, however, at R9 380 ITX - almost if not exactly the same format. And with a full 256 bit memory interface (while I doubt that the 3.5+0.5 config would change anything wrt to memory lanes on the PCB).

 

[ImSpartacus]

That's right. And although the 970 may have a somewhat simplified memory footprint due to the 224bit bus, I doubt that taking away 1/8th of the connections made a huge difference.
Then again, we never saw a mini 980.. There's a "980 for laptops" but I don't know how big it is.

Would the layout of a 970 be different due to that infamous memory quirk?

From my limited understanding, it still had all of the dram (and consequently traces?) that a 980 would have. Therefore, I had interpreted it as more of a 224+32 bit bus.

Anandtech did a good write-up (as usual).

 

[gamervivek]

He's talking of C4 version despite its clockspeed being the same as the C7. So it's more likely that it's some other difference, drivers perhaps.

Comparisons of the single card results.

http://www.3dmark.com/compare/3dm11...3dm11/11263252/3dm11/11167887/3dm11/11167781#

The CF scaling is bad but then it's a 720p benchmark.

 

[Love_In_Rio]

http://videocardz.com/60253/amd-radeon-r9-480-3dmark11-benchmarks

C7 is 2304 shader cores by SiSoft.If true it would be the R9 480 Pro, not XT.XT with higher clock would be easily faster than Fury X.

 

[Deleted member 13524]

C7 is 2304 shader cores by SiSoft.If true it would be the R9 480 Pro, not XT.XT with higher clock would be easily faster than Fury X.

I'm still skeptical of this. Polaris 10 would've made a lot more sense as a laptop-friendly product that gives Hawaii performance at ~100W.
With that approach, AMD could still be making Fiji cards and selling them at for example $300-350 until Vega comes out. This way, all Fiji cards will be cannibalized overnight.

 

[silent_guy]

From my limited understanding, it still had all of the dram (and consequently traces?) that a 980 would have. Therefore, I had interpreted it as more of a 224+32 bit bus.

That's exactly right.

 

[3dilettante]

I'm still skeptical of this. Polaris 10 would've made a lot more sense as a laptop-friendly product that gives Hawaii performance at ~100W.

That provides quite a bit of leeway for it to be driven harder. Making up a deficit with the Fury non-X might be possible.

With that approach, AMD could still be making Fiji cards and selling them at for example $300-350 until Vega comes out. This way, all Fiji cards will be cannibalized overnight.

Is this a bad thing?
Fury has been kept out of that cheaper and higher-volume niche so far. Maybe it's not that much of a loss if that's the only place it can go.

 

[Ext3h]

Is this a bad thing?

Yes, it is. The rest of the lineup can be relabeled and sold of to OEM partners at a heavily discounted price.

Doesn't work with Fiji. Too expensive to clear the stocks like that, there is a pretty high lower border AMD can't undercut.

 

[Razor1]

wonder if they will take a write off of remaining fiji parts.......

 

[Love_In_Rio]

I'm still skeptical of this. Polaris 10 would've made a lot more sense as a laptop-friendly product that gives Hawaii performance at ~100W.
With that approach, AMD could still be making Fiji cards and selling them at for example $300-350 until Vega comes out. This way, all Fiji cards will be cannibalized overnight.

Isnt better for AMD to have a killer architecture that can reach several segments with one chip?.

 

[Razor1]

Left over inventory is never a good thing to have around.

 

[3dilettante]

Yes, it is. The rest of the lineup can be relabeled and sold of to OEM partners at a heavily discounted price.

Doesn't work with Fiji. Too expensive to clear the stocks like that, there is a pretty high lower border AMD can't undercut.

Giving it that kind of breathing room means hobbling Polaris, which should be a smaller more cost-effective design that is gauged against long-term competition from Nvidia's Pascal generation.

In terms of die size, even with the more expensive process, Polaris 10 is rumored to win. In bandwidth it might falter, but at the same time the choice in memory might be salvageable given Fury's limited ability to leverage HBM and the measurable yield hit taken with HBM and interposer assembly.

AMD's managed to burn some of that Fury stock by putting dead Fury packages in some of its Radeon Pro Duo kits.

wonder if they will take a write off of remaining fiji parts.......

I was going to quip that AMD has plenty of experience taking charges due to inventory control failures.

They could put Fury chips into Duo, for a little while, perhaps. The pricing should have kept the expected volumes in a lower tier than a 300-dollar card would have, so hopefully AMD hasn't accumulated too much of it.

On a side note, Nvidia is rumored to have cut off 980 Ti manufacturing, which is the upper bound on how salable Fury is and should give a picture of how much room is left for a Fury rebrand to burn inventory.
It might suck more to build up 14/16nm inventory or put more downward pressure on its pricing just to give Fury shelf space.

 

[Jawed]

One way to use up the inventory, Ethereum mining:

 

[3dilettante]

Why not at this point?
AMD could create a slightly cheaper mining edition of Fury inventory with the display outputs ripped out, or even give it one lonely restricted/outdated (put in a converter for VGA) output so that the graphics market is not hit by older GPUs in 6 months that have their resale price subsidized by mining.

If that makes them uncompelling for the GPU miner of the week, then AMD can just write it down and get it over with.

 

[Esrever]

Why wouldn't the miners want the new energy efficient cards at that point?

 

[Razor1]

Its going to come down to how close Polaris can get to Fury with arithmetical levels, right now its rumored around ~5.5 tflops, which is quite a bit less than Fiji.

 

[Esrever]

Well If it only uses half the power, then your profit from mining would be much higher if you are paying for electricity.

 

[Grall]

One way to use up the inventory, Ethereum mining:

Wait wuzzat?

...And jesus, 30 fury nanos? Someone's made a sizeable investment! That's like...half of AMD's total Nano production run?

 

[3dilettante]

Well If it only uses half the power, then your profit from mining would be much higher if you are paying for electricity.

The Fury Nano was given a typical board power of 175W. Which wattage level is being assumed for Polaris 10 for this comparison? None that have been discussed do better than maybe 2/3 the power, which is close to where the rumored throughput ration would be.