RE: Inquiries on the Axon hardware

owadekimo — Wed, 19 Aug 2026 14:55:42 GMT

Hi,

Thank you for the clear and detailed response. That the Axon NPU is pipelined in the way you describe makes a lot of sense in retrospect, and certainly clears up our benchmarking efforts.

RE: Inquiries on the Axon hardware

SwRa — Wed, 15 Jul 2026 14:18:30 GMT

Hello,

Sorry for the delay.

The Axon NPU TCM is very small relative to the model size, it is essentially a cache. Axon NPU is continuously reading from/writing to system memory, to/from TCM. So rather than reading in an entire layer, Axon NPU loads a small portion, consumes it in its entirety, then reads in the next portion. This is all pipe-lined so there are minimal stalls for waiting for data. The layer input/output activations are always read from/written to the interlayer buffer. In some cases there are intermediate results within a layer that need to be stored to it as well. With convolutions this is not the case though.
So you are mostly correct, it's just that the job is performed in parallel with reading in the input and weights and quantization parameters into its cache and performing compute on the cached values.

Best Regards,

Swathy

RE: Inquiries on the Axon hardware

SwRa — Fri, 03 Jul 2026 13:16:13 GMT

Hello,

Let me look into this and get back to you.

Regards,

Swathy