For all the people saying that AMD is still competitive: look at the GPU sizes. Traditionally, AMD was able to compete with Nvidia using similar die sizes, sometimes even smaller. That's very important from a business perspective, because larger dice cost more money, use more power, and usually also have wider memory buses (which increases the cost of the PCB and auxiliary components).
Against Kepler, AMD was pretty competitive. Tahiti (352 sq.mm., 4.3 billion transistors) was a bit bigger and more power-hungry than GK104 (294 sq.mm., 3.5 billion transistors), but it wasn't a huge difference and was partially justified by Tahiti's better compute performance (especially in Double Precision). These chips were at least in the same basic class. On the extra-large dice, AMD came out ahead; 438 sq.mm. Hawaii (6.2 billion transistors) was a better gaming chip than the 561 sq.mm. GK110 (7.1 billion transistors) and was competitive in GPGPU (including Double Precision performance) as well. That said, even though AMD had the technical advantage on this battle, Nvidia won the business competition by a mile; they made far more on GK110 Tesla than AMD could dream of doing on FirePro Hawaii, and on the consumer side, they first milked the high-end market for $999 Titan sales, and then beat AMD on the midrange in large part simply by providing a stock cooler that didn't totally suck. AMD really botched the Hawaii release badly, and it hurt them a lot. IMO, it should have been a pure "virtual" release (no reference cards), and the R9 290 should have had a lower power limit for more energy efficiency since unlike the R9 290X it didn't need to win the performance crown outright. Another example of AMD's competent engineers being let down by their abysmal marketing team.
Against Maxwell, though... AMD simply can't compete at all if you look at what kind of chips they have to price against Nvidia's offerings to retain even what little market share they have. Consider the following: Pitcairn is a 212mm^2 chip with 2.8 billion transistors. GM206 is a 227mm^2 chip with 2.9 billion transistors. These chips should be in roughly the same performance and price bracket - but they're not even close, since AMD is still using an obsolete 2012 design. AMD is lucky to get $130-$150 for a Pitcairn card, while Nvidia's GM206-based GTX 960 is selling like hotcakes at $200-$240 (depending on RAM, featureset, etc.) And it's even worse than that, because the GM206 only has a 128-bit bus while Pitcairn has a 256-bit bus. So the AMD boards are more expensive to produce: more traces and more RAM chips are required. To compete with GM206 in terms of performance, AMD has to use Tonga - a much larger chip (359 mm^2, 5.0 billion transistors) with higher power requirements. They therefore sell fewer cards at the same price point, and make less profit on those they do sell. And that means fewer R&D dollars for the next generation (if it ever comes).
Likewise, we've heard people say that Hawaii can compete with GM204 (GTX 970/GTX 980) in terms of performance. Sure, if the only thing you're concerned about is raw framerates in today's AAA games, it can. But in every other metric (except Double Precision, where it still competes with Kepler), Hawaii falls way behind. GM204 is a 398 sq.mm. chip with 5.2 billion transistors: not much bigger than Tonga. Yet the Tonga-based R9 285 has less than 65% of the GM204-based GTX 970's performance. To get that performance level, AMD has to resort to Hawaii, with a 438 sq.mm. chip that contains 6.2 billion transistors - and they have to push it far beyond its optimal performance per watt, which hurts sales. Hawaii has a massive 512-bit memory bus, and it needs a very robust power stage, which makes boards expensive to design. Profits on Hawaii must be tiny compared to profits on GM204. Again, this means less money to sustain AMD's development for newer graphics chips.
Facebook
Twitter