[libre-riscv-dev] AI Accelerator

[Immanuel, Yehowshua U]

> Turing completeness bites again…

Hahaha

Luke does have a point though.

Luke, our accelerator specifically targets accelerating CNNs and NLPs.
It does this by having a hardware binary adder tree with reconfigurable forwarding.
We only execute at 8-bit precision since it turns out that well trained CNNs and NLPs are highly
resistant to quantization…

That’s how we get such great speedups. This does come with some drawbacks however.
Namely, the accelerator can ONLY do inference - not training.

This is where NVIDIA currently shines. Their Volta architecture is the fastest AI training architecture in existence at the moment.

So while image classification and language processing are very intelligent behaviors, machines with general intelligence and task solving capabilities don’t really exist yet. So terminator is quite a while away.

The application of an AI accelerator in this sense might be a drone that can do improved path planning because it can quickly classify images at native FPS.

Or perhaps some augmented reality glasses that can pin information to objects in the frame - again because of image classification.

Or lastly, a device using a LibreSOC that transcribes your speech or translates it. All these tasks of image classification and NLPs are used done ubiquitously across a large class of consumer mobile hardware today.

I think its a disservice to “true intelligence” to call it AI - but then again, its so much better than what existed 10 years ago.

Its good we’re having this conversation. Its making me consider some important points.

Yehowshua