The iPhones XS • Hacker News Board

Re: The iPhones XS

#31

Post by kartickv » Fri, Sep 21, 2018, 5:54 AM UTC

ValentineC wrote: post #13

> … to play with the bokeh depth of field f-stops, you have to use an iPhone XS because it depends on the A12’s Neural Engine. An iMac Pro has a more powerful CPU than an iPhone XS, but it doesn’t have a Neural Engine, and the bokeh effect depends upon it. This doesn't bode well for anyone who cares about being able to manipulate their photos without being too tied into Apple's ecosystem. I hope that Adobe and the de…

Agreed. Not just a Mac but any PC should be able to adjust the aperture using software. Let it be slow. Better than not being able to do it at all.

Re: The iPhones XS

#32

Post by Steko » Fri, Sep 21, 2018, 10:59 PM UTC

kartickv wrote: post #30

Steko wrote: post #27

Earlier quoted context omitted.

These chips, like GPUs, are optimized for parallel processing, but they are also designed for lower precision math which requires less memory and fewer transistors/operation and hence are considerably more power efficient.

So it's just a GPU optimised for half-floats and/or integers? I remember reading that it's optimised for matrix calculations. Is that true? If so, what is it about matrix calculations that makes a GPU not the most optimised for this task?

Yes it's optimized for 8 bit matrix multiplication.

> If so, what is it about matrix calculations that makes a GPU not the most optimised for this task?

Well this gets into what is a GPU since now NVIDIA and others are making special tensor GPUs that are also good at this. Most GPUs are optimized for 32 floating point math though, you're probably familiar with "8-bit graphics" being synonymous with 80's NES console quality.

Re: The iPhones XS

#33

Post by kartickv » Sat, Sep 22, 2018, 2:57 PM UTC

Steko wrote: post #32

kartickv wrote: post #30

Earlier quoted context omitted.

So it's just a GPU optimised for half-floats and/or integers? I remember reading that it's optimised for matrix calculations. Is that true? If so, what is it about matrix calculations that makes a GPU not the most optimised for this task?

Yes it's optimized for 8 bit matrix multiplication. > If so, what is it about matrix calculations that makes a GPU not the most optimised for this task? Well this gets into what is a GPU since now NVIDIA and others are making special tensor GPUs that are also good at this. Most GPUs are optimized for 32 floating point math though, you're probably familiar with "8-bit graphics" being synonymous with 80's NES console q…

I see. When I said GPU, I meant a normal GPU, not a special tensor GPU. BTW, what makes tensors / matrices hard for GPUs to handle?

Re: The iPhones XS

#34

Post by Steko » Sun, Sep 23, 2018, 7:01 AM UTC

kartickv wrote: post #33

Steko wrote: post #32

Earlier quoted context omitted.

Yes it's optimized for 8 bit matrix multiplication. > If so, what is it about matrix calculations that makes a GPU not the most optimised for this task? Well this gets into what is a GPU since now NVIDIA and others are making special tensor GPUs that are also good at this. Most GPUs are optimized for 32 floating point math though, you're probably familiar with "8-bit graphics" being synonymous with 80's NES console q…

I see. When I said GPU, I meant a normal GPU, not a special tensor GPU. BTW, what makes tensors / matrices hard for GPUs to handle?

AFAIK GPUs are ok at matrix math, it's just they are typically built for higher precision. AI chips probably have some other advantages but I'm not the best person to speak to them. Here's my simplified understanding:

CPUs are optimized for taking a decent amount of data and doing anything with them, especially if the order counts as is often the case with algorithms.

GPUs are optimized for taking a lot of data and doing the same few things to a lot of them (shaders, physics).

AI chips are optimized at taking a massive amount of data and doing a large number of calculations on all of them. In a neural net each "neuron" and/or "synapse" can represent a function/weight that needs to be calculated across many inputs.

In a deep learning set, this could be millions of weights and the more weights you have the more important memory and power efficiency become. When a NN is being looked up into (as opposed to the initial training) it's often in real time (self driving cars, "hey siri", etc.) so the demand to finish all those calculation's faster is even higher. OTOH the accuracy is less important, as neural nets - like people - are good at filtering out noise.