If these are going to be cheap it would be a game changer for the hobbyists/ hacker market, get the M.2 version and an SBC or a board for this and the Pi compute module run your main logic on the ARM CPU and use this the Mythic chip as an accelerator for object recognition or localization.
Now it will all boil down to software and how easy would it be to use these darn things, if they are going to be cheap enough it doesn’t even matter if they’ll become obsolete quickly since you’ll be able to just swap them.
Flash has limited write cycles, I wonder what the lifetime of these chips are supposed to be. I would think trying to compute with them would hit that limit very quickly.
The weights are loaded once per model, so you would only have a write when you update the model, which happens once every few hours or days for most applications today.
An interesting notion that machines with these will experience a form of dementia as their memory cells fail to write and start reading as old or corrupt data.
Fortunately, we can always move such a mind to a brand-new brain, able to hold more data for longer. Or simply keep it synchronized all the time, so that respawning would be immediate.
Potentially just a few of those cores alone will allow for wakeword recognition! This in-memory analog compute breakthrough is to me, in the truest sense of the word, mindblowing. I've been following Mythic since learning of memristor-based tech for TinyML applications and I'm beyond extremely impressed with not just the engineering but also the market interaction side of it.
So this is purely for inference right? Basically you're setting up a physical analog inference device by writing weights to flash memory and then using the device properties to do that mat-muls required for the forward pass?
This is super cool, I'd really like to see how this shakes out in actual products in terms of models supported, actual costs and performance, etc.
10 comments
[ 3.1 ms ] story [ 31.6 ms ] threadNow it will all boil down to software and how easy would it be to use these darn things, if they are going to be cheap enough it doesn’t even matter if they’ll become obsolete quickly since you’ll be able to just swap them.
Consumable chips I guess.
Fortunately, we can always move such a mind to a brand-new brain, able to hold more data for longer. Or simply keep it synchronized all the time, so that respawning would be immediate.
This is super cool, I'd really like to see how this shakes out in actual products in terms of models supported, actual costs and performance, etc.