[lnsru]

Working in the area makes this text on Tensil’s page me extremely curious: “This gives you the performance of custom silicon at the cost of commodity hardware. With Tensil, you can roll out your dream chip in weeks instead of years, at prices measured in thousands instead of millions.” So they designed some fuse programmable ASIC for machine learning? Setting the fuses makes this ASIC “custom silicon”? ASIC for thousands? Sounds to good to be true.

[323454]

Tensil founder here. We're a little different from the other ASIC companies because we specialize our chips to individual model architectures on demand. Our value prop is that our chips are smaller and cheaper than high end GPUs while offering the same or better performance on the model you care about.

This is compelling to a lot of hardware companies, where they often have 1 model that represents >50% of their compute workload. Making a Tensil ASIC for that model becomes much cheaper than GPUs at scales of 1000 chips or more.

I'm happy to talk in more detail if this interests you! [e-mail]

[LeonM]

Is there even a market for cheap ASICs?

ASIC set-up costs are only a problem on low volume production runs, and in that case generic purpose hardware is probably a cheaper solution anyway.

If you have some high computation or bandwidth requirements that can't be met with of the shelf hardware, then FPGAs are probably still a better solution.

So if you ask me, Tensil is just trying to ride the AI wave.