Not sure why you're saying that systolic arrays never went anywhere. They're widely used for matrix operations in many linear algebra accelerators and tensor units (yes, largely in research), but they are literally the core of Google's TPU [1] and AWS EC2 Inf1 instances [2].
[1] https://cloud.google.com/blog/products/ai-machine-learning/a...
[2] https://aws.amazon.com/blogs/aws/amazon-ecs-now-supports-ec2...