![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
Nvidia increased their boost from ~1100 to 1700 MHz from Maxwell to Pascal. Even if we use 1200 MHz (a more realistic number) to 1700 MHz is still a massive improvement.
that's a terrible way to compare man, we know nvidia was very conservative with clocks on maxwell... we have to wait and see how much headroom (in %) is available on gp104, they skimped on power efficiency; they chose a higher clocked, smaller, less efficient die over a larger, more efficient, lower clocked die.
At this point I would believe anything up to ~1550 MHz sheet boost is possible.
GTX 1080 is 180W and let's say 25% faster than the Titan X.
In other words taintedsquirrel, nvidia and amd had their 28nm dies manufactured on the same process, why were they competing with 400mm gpus with 540mm gpus ? 220mm vs 340mm etc etc
True but at this point I sort of expect a clock speed bump of some sort. To shrink the chip in half and keep the clock the same and not squeeze anything more is pretty lame. I mean we had R9 series hitting closer to 1150 core easily but it wasn't much more stable after that on air. I do expect a few hundred mhz atleast. Damn am I asking for too much here, lol.
Are they? VR is well suited for multi-adapter. Getting 2 chips consuming <150W onto a board capable of handling max 300W for $400-500 would definitely bring VR to the mainstream. $1k and you could have a VR headset with a dual Polaris 10. That gives them a $200 entry level VR card and an easy option to double performance without waiting for Vega. One other consideration, cards may need 100W headroom to power a thunderbolt interface for a VR headset. The headset shouldn't require that much, but that's the spec. I'd have to check on how power passthrough was handled,, but the cards may be responsible for powering their displays in the future to reduce cabling.
That's only accounting for clockspeeds. If actual utilization increases as suggested by those patents/papers, say 30% based on benchmarks in the paper, they'd only need ~1200MHz to reach Fiji levels. This would be ACTUAL performance, not necessarily theoretical performance.
From just the process no, but that doesn't mean the CHIP can't double clocks, cut the die in half, and reduce power.
If AMD can at least manage to match the 980 Ti / Titan X I would have no problem sidegrading. I think your prices are high though, $400 for a ~230mm chip, AMD will be rich.
They should be less than half the size. The $200 part is likely what the consoles are getting with a cut down Polaris 10, relatively slow memory, and low TDP to benefit power circuitry and the cooler costs.
May not cover the actual changes, but all the features the paper lays out as being possible through their technique line up with the GCN improvements AMD included on a slide. They are all enabled by the same capability. One of the authors becoming an AMD Corporate Fellow less than a year later probably isn't a coincidence either. Granted he was an old ATI guy, but still.
Well they'd need compute bound workloads to really shine and probably the ability to gate off ALUs to reduce power draw to what is required.