Back in May, Bloomberg's Mark Gurman broke the news that Apple (AAPL 0.51%) is working on a specialized processor "devoted specifically to AI-related tasks."
Late last month, I suggested that the Apple Neural Engine wasn't a stand-alone chip. Instead, it sounded more like a processor that would be integrated into Apple's A-series applications processor.
Image source: Apple.
After all, Apple's A-series processors is what is known as a "system-on-a-chip," or a chip that integrates many different pieces of functionality. For example, the A-series processors today integrate a whole host of technologies such as CPU cores, graphics processors, memory controllers, image processors, and even a motion co-processor.
However, based on what Gurman said in a recent article, it sounds as though the upcoming Apple Neural Engine is, in fact, a stand-alone chip.
The evidence of a stand-alone chip
"In addition to the face unlock feature, Apple is testing next-generation iPhone prototypes that include a dedicated chip for processing artificial tasks," Gurman said.
At this stage of the game, the Apple A11 Fusion system-on-a-chip should already be in mass production, and even if mass production hasn't yet commenced, the design must have been finalized quite a long time ago.
If the Apple Neural Engine is integrated into the A11 Fusion, then I wouldn't expect Apple to be "testing" anything -- it'd be good to go for the soon-to-be-launched iPhone models.
This wording, then, leads me to believe that the Apple Neural Engine is, in fact, a stand-alone chip that's separate from the A11 Fusion applications processor.
Alternatively (though, I think less likely), the Apple Neural Engine might be included in the A11 Fusion chip, but Apple is now deciding if it wants to enable it.
A sensible move, but the future is still integrated
If the Apple Neural Engine was a relatively risky/ambitious piece of silicon, then it could make sense that Apple wouldn't want the development schedule of the A11 Fusion chip -- a chip that's critical to the functionality of the new iPhone -- to be put at risk due to the Apple Neural Engine.
So, for a first iteration of this new, unproven, and -- at least, right now, inessential -- technology, building it as a separate chip is sensible.
However, because stand-alone chips take up precious logic board space and can lead to additional cost and power consumption overhead relative to technologies integrated into the main processor, I wouldn't expect the Apple Neural Engine -- if it really is a stand-alone chip -- to remain a stand-alone chip forever.
Think, for example, of Apple's M-series motion co-processors. The first M-series motion co-processor, known as the M7, was a stand-alone chip, and so was its successor, the M8. However, the M9, M10, and presumably all future M-series co-processors are all integrated directly into the A-series silicon.
Indeed, it wouldn't surprise me a bit to see future iterations of the Apple Neural Engine (there's no way Apple doesn't improve it, and perhaps dramatically so, in the coming years) integrated directly into the A-series processors. Perhaps such integration could come as soon as next year.
