5 ms·
Faster, probably, but 2x+ Faster like in PowerPC > Intel very unlikely. Which is why there’s much less likely to be enough performance budget for effective emul
by drham 8y ago
Faster, probably, but 2x+ Faster like in PowerPC > Intel very unlikely. Which is why there’s much less likely to be enough performance budget for effective emulation.
- saagarjha 8y agoPresumably their "desktop-grade" A-whatever chip will have a higher TDP and clock speed than their mobile counterparts.
- gascan 8y agoIf they could just turn up the TDP and clock speed on their chips and get 2x the performance of Intel's best chips that easily- they would already own the entire desktop market.
- mozumder 8y agoAnd some of us, with actual CPU design backgrounds, have been saying that for a while now. We've been asking why Apple doesn't already do it.
- samcat116 8y agoAs someone without a CPU design backround, would something like this scale pretty linearly with added power and thermal headroom? I assume there are limit that would have to be overcome, but what would an ARM chip in the conditions of an i7 look like?
- mozumder 8y agoIt's more of a logarithmic scale. The same architecture can generally scale to 10x over a few process generations.
- arbie 8y agoThis does not hold at 10nm and below.
- bzbarsky 8y agoTDP typically scales as somewhere between the cube and fourth power of the clock speed if you're pushing the envelope (in the sense of running at frequencies where further frequency increase also needs a voltage increase). So having 10x the thermal envelope means you can probably clock about twice as fast, all else being equal.
- BillinghamJ 8y agoSeems that it’s more of a logistical and network-effect issue in getting everyone to support it, rather than a technical issue. Possibly a patent issue also
- mozumder 8y agoMy only question is if they actually use ARM for their next-generation architecture, or something completely new...
- mozumder 8y agoComputer architectures routinely see 3x performance jumps across different power budgets. This rule has held over decades. Clock speeds alone can probably increase by 30%. Caches and internal datapaths can double or more. Then you can start to add in more execution units or more expensive branch-prediciton or even new more power-hungry instructions. A 4 Watt Intel Pentium 4410Y Kaby Lake for mobile devices gets about 1800 on Geekbench, while a 115 Watt Intel Core i7-7700K Kaby Lake for desktops gets 5600. I'm just going to say it: the Apple laptop CPU is going to get Geekbench score... above 9000! And, yes, I do have a CPU design background.
- michaelmrose 8y agoSo artificial benchmarks already do a very poor job of capturing performance. The apple laptop cpu does not exist. If it did exist it would likely suffer a very substantial performance hit if forced to emulate x86 software. So you are going to speculate on the meaningless benchmark numbers of an imaginary cpu that will take a wholly unknown hit if everyone does not rewrite everything why?
- mozumder 8y agoArtificial benchmarks do a great job of capturing performance, since they're more controlled and eliminate unnecessary variables. Once you understand this, then you can understand how CPU designers work to predict future performance. CPU designers use artificial testbenches.
- michaelmrose 8y agoYou making up numbers doesn't appear to be a useful endeavor.
- sjwright 8y agoI suspect if Apple designs a desktop CPU, performant x86 emulation will be a key design criteria. I know very little about CPU design, but I imagine it would be possible to have hardware optimisations for x86 emulation just like we have today for video codecs. Or even further they could bake a "rosetta" into the chip's microcode and have their CPU natively support the x86 instruction set along with ARM or whatever they come up with.
- megy 8y agoDo you think they will use the same chip? In 2020?