Benchmark Comparison

EyeQ™6H vs. NVIDIA Jetson AGX Orin

EyeQ™6H vs.
NVIDIA Jetson AGX Orin

EyeQ™6H delivers high-performance AI at the edge, deploying DNNs, transformer models, and ML frameworks for real-time classification, object detection, and segmentation. Our most advanced SoC yet is designed to transform raw sensor data into precise, actionable insights.

Image Classification

TOPS

34 TOPS

275 TOPS

Single Stream Latency (msec)

Mode

Power

Power (MaxQ)

Unconstrained (MAXN)

ResNet-50

0.5 msec

1.64 msec

0.64 msec

Power mode refers to testing a chip in its original state without any modifications.
Unconstrained mode refers to testing without any artificial power or thermal restrictions.

* Reported by Nvidia and MLPerf.

Vision Transformer

Mode

Power

Unconstrained

EfficientViT-B1

224x224

9.1M

0.52G

0.564 msec

1.48 msec

EfficientViT-B2

224x224

24M

1.6G

0.932 msec

2.63 msec

* EyeQ™6H is designed for Int8 computation, to optimize performance, efficiency and accuracy by leveraging advanced QAT (quantization aware training) and PTQ (post training quantization) techniques.

** Reported results for Orin are in FP16.

Image Classification

Power mode refers to testing a chip in its original state without any modifications.
Unconstrained mode refers to testing without any artificial power or thermal restrictions.

* Reported by Nvidia and MLPerf.

Vision Transformer

** Reported results for Orin are in FP16.

Architecture of Efficiency

With a unique and highly efficient architecture of diversified accelerators, EyeQ™ achieves state-of-the-art computer vision performance within a low-power envelope

Heterogeneous Computing

Using the most suitable core for each task

From general-purpose CPU cores to high compute density accelerators, including deep learning neural networks

CPU

Central Processing Unit

MPC

Multi-threaded Processor Cluster

More versatile than a GPU, and with higher efficiency than any CPU

VMP

Vector Microcode Processor

A wide vector (VLIW and SIMD) machine with exceptional performance for short integral types common in computer vision and deep learning algorithms.

PMA

Programmable Macro Array

A CGRA dataflow machine. Its unique architecture delivers outstanding performance for dense computer vision and deep learning algorithms that are unachievable in classic DSP architecture.

XNN

Deep Learning Accelerator

Dedicated high-performance AI engine. The main source of horse power for convolutional neural networks.

GENERAL-PURPOSE COMPUTE

CPU