[Anduia]

> Critically, this processor achieves accuracies approaching those of conventional 32-bit floating-point digital systems “out-of-the-box,” without relying on advanced methods such as fine-tuning or quantization-aware training.

Hmm... what? So it is not accurate?

[btilly]

It's an analog system. Which means that accuracy is naturally limited.

However a single analog math operation requires the same energy as a single bit flip in a digital computer. And it takes a lot of bit flips to do a single floating point operation. So a digital calculation can be approximated with far less energy and hardware. And neural nets don't need digital precision to produce useful results.

[B1FF_PSUVM]

> neural nets don't need digital precision to produce useful results.

The point - as shown by the original implementation...

[bee_rider]

It seems weirdly backwards. They don’t do techniques like quantization aware tuning to increase the accuracy of the coprocessor, right? (I mean that’s nonsense). They use those techniques, to allow them to use less accurate coprocessors, I thought.

I think they are just saying the coprocessor is pretty accurate, so they don’t need to use these advanced techniques.