Intel Corp (INTC)

[wbmw]

The plot thickens on Apple A8

3DMark Icestorm submission: 17,838

http://www.futuremark.com/hardware/mobile/Apple+iPhone+6+Plus/review

I was trying to think of a good comparison point. Larger form factors tend to give additional thermal headroom, though the iPhone 6 Plus is a fairly large design. But here's a couple of comparisons which might help triangulate gen over gen performance:

iPhone 6 Plus vs. iPad Air: 17,838 vs. 15,183 (17% improvement)
iPhone 6 vs. iPhone 5S: 17,284 vs. 14,702 (18% improvement)

Either way, we're talking improvements far below what we saw in GfxBench. And though I'm sure the response from some people will be that GfxBench is the "only benchmark that matters" or "the only benchmark that's popular" (and we can all debate those) - this is a second datapoint, suggesting that Apple's new 20nm part may deliver substantially less performance in other benchmarks than what they quoted at launch.

And, mysteriously, it took them "~2B" transistors to rev the design. Although this was argued as one of the more impressive feats of engineering (mostly due to the density improvements in transistors per mm2) - it doesn't give Apple a lot of headroom if they want to improve the design going forward.

If we assume that there won't be any density improvements in the near term, there's nowhere to go, but to grow the design. The 16nm FinFET process will improve power/performance, but Apple doesn't have a power constraint - they are moving towards a transistor constraint. Doubling the size of the graphics (based on the Chipworks teardown) will grow the die another 20mm2. Moving to quad core (based on the same teardown) will grow the die another 15mm2. And that doesn't include any additional growth in memory bandwidth, or any other on-die functionality. If we assume another 15mm2 of additional growth from other units, then the A9 very easily grows to 140mm2.

Meanwhile, Broadwell is 82mm2, and already delivers nearly 3x the performance of A8 in 3DMark Icestorm, so it should continue to perform better against A9 as well. I'll also go out on a limb, and predict that Broadwell gets close to the GfxBench performance of 15W Haswell, which is at 30 FPS, while the A8 gets 19 FPS. Based on the same assumptions before, Apple doubles their die area in graphics and builds a 140mm2 design for A9 on 16nm FinFET. By the time you figure in Intel's incremental speed bins for Broadwell, I think Apple A9 will be at best about even with it, in the most favorable Apple benchmark. They'll still be behind in Icestorm.

And that's for a 140mm2 projected die size, due to the fact that 16nm FinFET from TSMC doesn't come with density improvements over 20nm. And this doesn't even account for the fact that Skylake will be shipping in the second half of next year, so Broadwell will be old news. Intel could really start to enable devices will a step-function of performance above Apple and other ARM devices.

Lots of unknowns still, but the data seems to be lining up.