- This topic has 1 reply, 1 voice, and was last updated 3 years, 10 months ago by
.
-
Posts
-
Hi
I have a multiplier-convolution algorithm that I would like to map onto 2803 bare metal or alternatively using a standard kernel method such as OpenCL. Looking through the documentation I find no mention of the internal architecture or programmers mode, and no SDK to support kernel or bare metal development. Could we start a discussion on when this will be made available and how it will look ? Without access to architecture and support tools for this then the 2803 is essentially locked from the programmer much like with Apple and IOS and so prevents the community exploring and developing on top of this very interesting and sensibly priced device.
In particular it is important for edge AI learning to be able run training algorithms and to optimise code on the bare metal to applications. For instance I have a new patented pipeline for highly efficient multiplier-less convolution at low precision that supports inference and training which likely map to the 2803 MAC array, but I do not know without the documentation and code to support this.
I made a couple of typos and the edit feature for the topic post is not working so oops there. The post should read <multipler-less convolution> as it operates using addition only and indexed addressing into sequentially computed 4-bit precision significand product terms with shift and add to the accumulator so performing coefficient sequential sparse convolution for CNN inference without the need for expensive multiplication.
- You must be logged in to reply to this topic.