[KongchangAI]
ConceptDeep Learning Accelerator

DLA

A fixed-function deep learning inference accelerator integrated by NVIDIA into Tegra/Orin series SoCs. It is designed around dedicated data paths for tensor operators such as convolution and can execute inference independently of the GPU, offering better energy efficiency than a general-purpose GPU but with limited flexibility.

Timeline (last 90 days)

Nov 13

From the perspective of architectural positioning, DLA's benefits are mainly reflected in energy efficiency and freeing up GPU resources rather than surpassing absolute throughput; planning task allocation by treating it as a GPU coprocessor is more in line with its design intent than expecting it to bear the main inference alone.

Unverified50%
Nov 13

NVIDIA Deep Learning Accelerator (DLA) is a fixed-function inference accelerator integrated by NVIDIA into Tegra/Orin series SoCs, specifically handling deep learning operators such as convolution.

Unverified50%
Nov 13

The DLA on Jetson AGX Orin provides multiple independently configurable accelerator instances, and each instance can host one inference pipeline.

Unverified50%
Nov 13

TensorRT can specify DLA as the execution backend when building an engine, and the builder determines whether each layer can execute on DLA based on the layer's operator type, precision and tensor shape, with layers that cannot execute or are not sufficiently beneficial falling back to the GPU.

Unverified50%
Nov 13

DLA is a fixed-function accelerator with a relatively limited set of supported operators and data types, so not all networks can be fully offloaded, and in actual deployment the network usually needs to be split into DLA-executable parts and GPU fallback parts.

Unverified50%
Nov 13

DLA has specific requirements for the memory layout of input and output tensors, commonly NHWC or specific blocked/aligned formats, while NumPy arrays and PyTorch tensors on the Python side usually adopt NCHW or contiguous row-major layouts, and when the two are inconsistent, reordering is needed before submission.

Unverified50%
Nov 13

DLA inference is usually submitted asynchronously and requires correct synchronization to read results, and the Python-layer wrapper should make this timing semantics explicit, otherwise problems of reading incomplete results are likely.

Unverified50%
Nov 10

The heterogeneous compute units of an in-vehicle SoC include CPU, GPU, NPU/DLA, and dedicated modules such as ISP and DSP; different units differ greatly in compute power, precision support, and energy efficiency, and placing the same model layer on different units can result in latency differences of several times.

Unverified50%

All Facts (8)

Unverified

From the perspective of architectural positioning, DLA's benefits are mainly reflected in energy efficiency and freeing up GPU resources rather than surpassing absolute throughput; planning task allocation by treating it as a GPU coprocessor is more in line with its design intent than expecting it to bear the main inference alone.

50%
Unverified

NVIDIA Deep Learning Accelerator (DLA) is a fixed-function inference accelerator integrated by NVIDIA into Tegra/Orin series SoCs, specifically handling deep learning operators such as convolution.

50%
Unverified

The DLA on Jetson AGX Orin provides multiple independently configurable accelerator instances, and each instance can host one inference pipeline.

50%
Unverified

TensorRT can specify DLA as the execution backend when building an engine, and the builder determines whether each layer can execute on DLA based on the layer's operator type, precision and tensor shape, with layers that cannot execute or are not sufficiently beneficial falling back to the GPU.

50%
Unverified

DLA is a fixed-function accelerator with a relatively limited set of supported operators and data types, so not all networks can be fully offloaded, and in actual deployment the network usually needs to be split into DLA-executable parts and GPU fallback parts.

50%
Unverified

DLA has specific requirements for the memory layout of input and output tensors, commonly NHWC or specific blocked/aligned formats, while NumPy arrays and PyTorch tensors on the Python side usually adopt NCHW or contiguous row-major layouts, and when the two are inconsistent, reordering is needed before submission.

50%
Unverified

DLA inference is usually submitted asynchronously and requires correct synchronization to read results, and the Python-layer wrapper should make this timing semantics explicit, otherwise problems of reading incomplete results are likely.

50%
Unverified

The heterogeneous compute units of an in-vehicle SoC include CPU, GPU, NPU/DLA, and dedicated modules such as ISP and DSP; different units differ greatly in compute power, precision support, and energy efficiency, and placing the same model layer on different units can result in latency differences of several times.

50%

Source Articles