0
\$\begingroup\$

We know that Apple's Arm A4 Chip that powered the iPhone 4 had approx 1.6 GFLOPs of processing power.

In the second generation of Apple's Airpods - there is a custom chip called the Apple H1.

My question is: Do we know the processing power (in GFLOPs) of Apple's H1 chip?

\$\endgroup\$
2
  • 3
    \$\begingroup\$ Does it even have a floating point unit? Most low power audio devices use fixed point DSPs to save power. \$\endgroup\$ Jul 12, 2020 at 3:17
  • 2
    \$\begingroup\$ I'd like to add my own personal experiences to @user1850479 's comment. In a product line from one of my clients, we chose to use the ADSP-21xx DSP which has no FP, at all, but includes things like a fully combinatorial barrel shifter -- which is darned useful if you are doing FP in software. It was very low power for the processing capability you got for it and nothing compared. It was the right choice. Later, Analog came out with their replacement, the BlackFin which does have FP. But watt for watt? It couldn't come close to competing with their decades earlier ADSP-21xx despite being newer. \$\endgroup\$
    – jonk
    Jul 12, 2020 at 3:42

1 Answer 1

1
\$\begingroup\$

Do we know the processing power (in GFLOPs) of Apple's H1 chip?

We don't, but I'll venture an estimate:

  • It needs to do bluetooth. You need no floating point for that.
  • It's not clear how much audio preprocessing it does, but let it do a bit of beamforming: that happens on the fixed-point audio samples; it might be advantageous to compute the (if used at all) singular value decomposition when estimating directions (I don't even think it does something like that), but that would happen maybe 10 times a second, and would mean maybe 1000 floatining point operations - by all means, you either do that in software, or you round your GFLOPS to 0 with an error of less than a millionth
  • According to the wikipedia page, it is sensitive to "hey siri" commands, but these are processed off-headphone. My alcatel OneTouch 511 was able to recognize spoken words (for speed dialing in the phone book), 2001. We can conclude that the necessary compute power was exactly 0 GFLOPS, with an error ranging in the parts per billion.

So, the answer is

0 GFLOPS.

This is not a device where having much floating-point compute power is an advantage. On the contrary, every transistor you flip, uses your precious battery energy; it's desirable for the device to do as little computation as possible.

\$\endgroup\$
2
  • \$\begingroup\$ hi! I'd love to hear why I've been downvoted! What is ill-described, or badly researched, or wrong, about my answer? \$\endgroup\$ Jul 12, 2020 at 12:40
  • 1
    \$\begingroup\$ +1 You can't win them all. I get downvotes on years old answers with tens of historical up-votes. \$\endgroup\$
    – Russell McMahon
    Jul 13, 2020 at 7:10

Not the answer you're looking for? Browse other questions tagged or ask your own question.