[usgroup]

I think you’re wrong about the ML precision. You need highly precise for most recursive machine learning tasks because you’re compounding errors otherwise.

Typically you can’t even use floating point representation: not accurate enough.

[SomaticPirate]

Disagree. https://arxiv.org/abs/1805.08691 demonstrates 8-bit architecture for a pre-trained CNN provides more than acceptable results with lower latency and higher throughput than a higher precision version.

[usgroup]

Sure but we were talking about precision rather than throughput ; tbh the result is hardly surprising .

[medius]

Analog can actually be much more precise than 8-bit since there's no quantization noise (yes, this is not A2D, but any intermediate results can take middle values that will only make the final value more precise than digital signal).