Tenstorrent Grayskull PCIe AI/ML Accelerators

Any idea how this would compare with Nvidia GPUs in price/performance for machine learning at home?

I guess this will be rather expensive (writing this from a Raptor Computing Power9 workstation so not much budget left) but it does look interesting:

> “Tenstorrent's processor cores are known as Tensix cores. Each Tensix core includes 5 RISC processors, an array math unit for tensor operations, a SIMD unit for vector operations, hardware for accelerating network packet operations and compression/decompression, and 1-2 megabytes of SRAM.”

They say that it works with pyTorch, ONNX and their own C++ API which sounds good. Over the years I accumulated a bunch of somewhat-exotic hardware that I could not actually use because of bad software support and documentation. Hopefully this will be better in that regard once it launches.

Tenstorrent needs a CEO like Stokes has in rocket engineering to tell its story or what Twitter is looking for next.