[frankinstien]

The lower power specs are impressive for these accelerators; .5 w per TOPS where it can perform at 4 TOPS, so that just 2 watts! My Vega 56 uses more than 200 watts and produces 21 TOPS in 16-bit mode. You can use several simultaneously,  so 6 would end up using 12 watts! But at a cost of $75 for the Intel NCS 2 and $130 for the Google CUA, 6 would end up totaling $450 for the Intel and $780 for the Google.

Has anyone on this forum used these yet?

[LOCKSUIT]

For a USB port? Or for AI? I'm lost bro...

[frankinstien]

For a USB port? Or for AI? I'm lost bro...

The Intel and Google accelerators are AI FPGAs and connect through a USB port.

[frankinstien]

I bought one, an Intel NCS2, from the samples it is faster when comparing to a CPU, but the sample code doesn't use all the horsepower of my server. The demo ran 3.5 fps with the CPU compared to 24 fps with the NCS2.  There is a voice recognition sample but my sound card, or should say the USB sound device since the machine uses a server board that has no built-in sound capabilities, would fail for whatever reason. I'm going to have to figure out what's going there. So the drawback is the NCS2 can't train models which means a GPU is still needed but once the model is trained the NCS2 can run it.

Anyway, it was only $75 and worth experimenting with.