Low-Cost High-Performance Through Sparsity

TensorDash is a deep learning component designed to deliver state-of-the-art performance with clever software-hardware co-design. This is achieved by exploiting sparsity during both training and inference on simple hardware design, using an innovative scheduler.

Low Cost

Our simple hardware design coupled with our innovative scheduler can exploit most of the potential with minimal hardware cost.

Memory Savings

The scheduler of TenosrDash works also as an on-chip compression engine that eliminates the zeros out of the tensors, boosting the effective capacity of the on-chip memory structures and significantly reducing the off-chip memory traffic.

Static and Dynamic

For inference on edge devices, TensorDash comes with our Tactical software scheduler that is co-designed with a hardware back-end to exploit model sparsity that is known a priori. TensorDash can also exploit dynamic value sparsity either in the activations during inference or in the activations, weights and gradients during training, using our innovative hardware scheduler.

Sparsity Structure Agnostic

Our design and scheduler do not rely on a specific sparsity structure. Whatever sparsity exists, our innovative offline scheduler will extract performance gains.

DataType Agnostic

Our technology is data type-agnostic whether it is a low precision fixed-point inference engine or a high precision floating point training accelerator, TensorDash boosts its performance and its effective on-chip memory and reduces its off-chip bandwidth consumption.



Sources of Sparsity

Weight sparsity can be induced during training by pruning unneeded connections. Activation sparsity can be boosted by activation function selection and clever quantization. Gradient sparsity can be increased by skipping inconsequential values in the backward pass. TensorDash can exploit all of them.

Our Services

Our expert team can provide the TensorDash design, integration and supporting software to provide custom solution to your design constraints and needs.

Our Solution

A simple low connectivity design coupled with a innovative scheduler that mimics a full crossbar design. We do more with less by leveraging a software-hardware co-design.

What You Need to Do

Almost nothing! TensorDash can be integrated on any data-parallel processing element.


©2020 by Tartan AI