We had an exchange in the comments earlier this week that seems to me like it's worthy of a little more attention. TR reader "ermo" asked me the following question:
Do you have any informed speculation to offer on whether we will see AMD update the R9 280X (GCN 1.0) and R9 290X (GCN 1.1) to Tonga (GCN 1.2) equivalent SKUs with 384 bit/3GB and 512 bit/4GB framebuffers in the upcoming R9 3xx refresh?
Which sort of performance improvements would you think we would be likely to see relative to the existing parts based on this speculation and extrapolating from your tests of the R9 285 2GB part?
Which is interesting! Now, he's asking for informed speculation, and that's really all I have to give at this point. Here are my thoughts:
My contention all along has been that Tonga has a 384-bit memory interface and more ROPs than we've seen in the R9 285. See the chip table in this review, and look at the die size and transistor count vs. Tahiti. Only makes sense.
Man, I wish we had a die shot to confirm it.
I think they've held back on the R9 285X or whatever due to existing inventory of Tahiti and Hawaii-based parts or chips. I think a full-fledged Tonga would offer performance well above Tahiti and not far from the R9 290.
The obvious logical move for the 300 series is Fiji at the top with 390 and 390X, Tonga in the middle with 380/X, and lower end stuff below that.
With the right pricing, AMD could be competitive on price and performance, but I don't think anything Tonga-based will match the GM204/GTX 980 on power-efficient performance. Even accounting for binning, the R9 285 draws too much power for that.
If Fiji is really a 20-nm chip, then it could match the GM200 on power efficiency, one would think, but only if it operates at conservative clocks. 20nm probably won't buy you much at higher voltage. The water cooling rumors around the 390X aren't making me optimistic.
Everything I hear about HBM is that it's expensive. That's my other big worry about 390X. If AMD can only match or slightly beat Nvidia with an expensive and exotic memory subsystem, that could make life difficult for them.
Still, I've been in the same room with a working 390X-based system, so they clearly have working silicon. That's a start.
That's what I think based on what we seem to know right now. Most of these things will either come to pass or be proven wrong in the next two to three months, so you can bookmark this post and come back later to taunt me if I got it wrong. At any rate, the 2015 GPU refresh cycle is already well underway, and things should get more interesting as AMD brings its new chip (or chips?) to market.