Wonder what Vega 20 is going to be. Navi is the 7nm GPU design now (Vega 20 seemed like it was planned to be a large GPGPU/compute focused card on 7nm, but seems like that has changed). I wonder if the idea isn't to start the road to mGPU Ryzen like GPU where they do an updated Vega that is smaller (i.e. goes to 2560-3xxx SPs, and keeps memory where it is, or would be great if they just add a third stack which should add some bandwidth), and so better suited for mGPU cards. Ditch the interposers and put 3 or 4 stacks of HBM between them. If nothing else, in games that don't do mGPU well, they could just leverage higher bandwidth and more memory from the shared pool.
Or maybe they're going to break the pro/compute/server chips from the consumer stuff like Nvidia did. Where the former gets larger chips with features/pieces tailored to those tasks. They could leverage Infinity Fabric on both.
AMD can make a profit per unit even at 300$ if they write off R&D expenses. It will be shit Polaris like margins, but they can do it. HBM2 and the die size aren't blocks for pricing.
If it's 699$, I imagine there's some reason for it.
Yeah. There's a price breakdown of Fiji and it cost ~$200 in parts cost (was less than $200 but figure with the HSF). Even if prices ballooned, it still should be able to break even on cost at $400 if not much much lower than that. Not that AMD would want to, but if they really want to sell cards they could (and then distribute R&D cost on APUs and pro/GPGPU). My guess is they'd rather sell them in higher margin pro markets until they can get costs under control and get software support better. I could see them putting prices high (getting mining buys, intentionally not wanting to get gamers til perf/$ is better), and then when they have to (impending Volta) or when it works better for them (maybe do package deals for like Black Friday or X-Mas, if they offer good prices with Ryzen bundles, especially say Threadripper - where the CPU margins can help offset the GPU; it would be an effective price drop on the CPUs as well without having to explicitly do that, and then they can have Ryzen+ or whatever the revision is, ready to keep sales going next year) they could offer deep discounts to get sales when production is better. And it'd give them time to do any minor rework or respins if that could help (although I'd expect that to be next year release, since I doubt there's any really minor things they could implement on the run).
Possibly, but its difficult to see what that would be based on what seems to be where Vega is at.
A cut down R9 Fury had a launch MSRP $100 less than the full-die, watercooled R9 Fury X. I don't see that changing to a $200 gap...
But I also am unsure about this rumored $699 price.
I potentially could see it, if the watercooled version is the only one they push to the full spec (and requires the AIO cooler to run it that way). If they offer a more efficient but lower performing air cooled one, that hits a much better perf/w (at the cost of outright performance), there would be a sizable performance difference plus the water cooler.
After the Hardocp test, I wonder if that isn't their plan. Make the case that they can offer similar power use at similar "smoothness" coupled with pushing the Freesync monitor price difference. Nvidia wouldn't be under a lot of pressure to drop price again on the 1080 if it has better absolute performance and much better perf/w, so it could let AMD get away with charging similar price with worse performance for a while. The watercooled version would get a switch, putting it in efficiency mode or performance mode.
Nano sized card at 250-300W? That's never going to happen.
I agree, but I could see them offering a Nano card (as it was popular).
There's supposed to be 2 chips, right? I could see them doing 4 cards. The top end pushed to the max watercooled one (that they only want diehard enthusiasts buying). A full chip but lower performing and clocked in a better efficiency spot air-cooled card (the one they want AIBs pushing to give it better coolers to handle it, and also higher price). Then a cutdown version of that card (that would offer closer to 1070 performance), and then below that a small Nano that is put at the clockspeed sweetspot, but that sits between 1070 and 580/1060 level in performance and price. Then later when they can bin more, the Nano and the cutdown card would just be Nano full chip cards, then AIBs offer the full sized custom cards, and then they let water cooled ones would be left based on stock (i.e. once they produce a certain amount that's it, if there's still some available they drop the price).