And that returns us to Polaris 10. What did AMD showcase? That's right, performance/watt because its what matters
Would it blow your mind if I told you that performance/watt can be multiplied by a target wattage (for point of example let's take the totally arbitrary value of 300W) and used to generate a really rough expectation of the performance of a part sized to use that much wattage? So if people have some unaccountable desire for a 300W part, they can get themselves hyped up over a demo that doesn't require a 300W part at all, and can be done with the smallest, best yielding chip.
Anyway, on the desktop processor front there's a pretty good reason why Intel's stagnating. Mobile can be dealt with great by performance/watt. Server's parallel enough that it can be dealt with by it because performance/watt means a core fits in a smaller wattage and server cores can be added to add performance, or alternately the chip can go full Xeon D and just cash that in on good performance that sips power and costs way less in the long run.
Mobile processors are continually getting a whole bunch of good enough for less and less wattage. Servers are continually getting more performance per watt no matter where on the curve they fall. So we know Intel isn't actually stagnating. So what gives? Why is the desktop not gaining like that? How much does the desktop benefit from extra cores outside serious enthusiast use cases? Not much at all. The problem is that the desktop is in the position where single thread performance really matters but it's got more capacity to dissipate wattage than can really meaningfully be used by current processor designs. So there's two options. First is clocking them up, but that's butting into an inability to clock the things much higher. Next up is trying to raise performance by using more transistors per core. That's all good and well but how wide can a core really meaningfully be? We're still trying to raise single threaded performance so we can't go full Power and just add more and more thread support to a core, so we're going to be chasing tiny shreds of performance with wads of transistors and those transistors use power. Well, that's a non-starter for mobile and server so congratulations you've just fragmented off a separate desktop uarch. Clean out your desk.
The nice thing is that increasingly the real power hungry enthusiasts can use more cores. So, for them you could in theory make a different platform that uses most of the work done for the server market where stuffing the thing full of cores is a good use of time and resources and generate a nice platform for higher core counts, and focus core count increases on that platform over time.
God I should've bought Haswell-E instead of Devil's Canyon.
Facebook
Twitter