tensor-native-processor.github.io

Latest News

Title

Accelerate tensor operations via tensor native processor

Writeups

URL

https://tensor-native-processor.github.io/

Summary

We are going to develop a tensor native processor and a compiler to accelerate tensor operations, such as tensor multiplication.

Background

We want to create a new tensor processor that is native to tensor operations. Compared to existing tensor processors, we want to make the following improvements to increase throughput:

The Challenge

This problem has the following challenges:

Resources

Goals and Deliverables

Platform Choice

Schedule

Week

Milestone Report

Schedule

Screen Shot 2022-11-29 at 5 04 20 PM

Summarization of work completed so far

We have written a native tensor processor in Verilog, a simulator that returns the number of cycles performed given a compiled program, and some sample programs performing operations on a single core.

How we are doing with respect to the goals and deliverables stated in our proposal

We are on schedule and believe we will be able to produce all “plan to achieve” deliverables. We should be able to produce a portion of “hope to achieve” deliverables by targeting the inference of a simple neural network (only involving Gemm and ReLU).

Things we plan to show at the poster session

Preliminary results

Operations Simulated Clock Cycles
ADD(v16, v16) 11
MULT(m16, m16) 122

Concerning issues