SetupAI
繁
SourceFreshcollected in 14h

Praxis One Brings 106 TOPS to Robots

Read original on Pandaily
#robotics#edge-ai#npu

A single-SoC platform targets heavy-duty robots with roughly 106 TOPS of local AI compute.

30-Second TL;DR

What Changed

Praxis One uses the MediaTek Genio Pro 5100

Why It Matters

Higher local inference capacity can reduce latency, cloud dependence, and connectivity requirements for robots. Developers will still need to validate thermal performance, model support, and real-time reliability in field conditions.

What To Do Next

Benchmark your robot’s perception and planning models on the Genio Pro 5100 platform, measuring latency, power, and thermal throttling.

Who should care:Developers & AI Engineers

Key Points

  • •Praxis One uses the MediaTek Genio Pro 5100
  • •The platform provides approximately 106 TOPS of NPU compute
  • •It targets heavy-duty robotics and embodied AI applications

Deep Insight

Background and context from public sources — not the original article. 7 sources cited.

Enhanced Key Takeaways

  • •Praxis One leverages TSMC's 3nm process node through MediaTek's Genio Pro 5100 SoC, incorporating an 8th-generation NPU (NPU 890) with a dedicated Transformer engine capable of running 7B parameter models at up to 23 tokens/second locally.
  • •The platform adopts an 'All-Big-Core' 8-core Arm v9.2 architecture (1× Cortex-X925, 3× Cortex-X4, 4× Cortex-A720) generating up to 260K DMIPS for robotic kinematics and path planning.
  • •Praxis One consolidates high-level cognitive decision-making ('cerebrum') and low-latency motion control ('cerebellum') onto a single SoC, replacing traditional multi-chip designs that separate cognitive accelerators from discrete motion MCUs.
  • •The hardware is engineered for industrial environments (−40°C to +105°C) and supports multi-sensor perception with up to 16 FHD camera virtual channels, hardware 16-bit HDR, and native Wi-Fi 7 / Bluetooth 6.0 connectivity.
  • •Huaruizhipu also introduced Praxis Nano based on the MediaTek Genio 720 to form a tiered embodied AI compute portfolio alongside Praxis One.

Competitor Analysis

Architecture
Huaruizhipu Praxis One
Single-SoC 'Dual-Brain' (Cognitive + Motion on one chip)
Traditional Robotic Edge Compute (e.g., NVIDIA Jetson / Discrete Setups)
Multi-chip split setup (Discrete GPU/Accelerator + MCU/x86 Host)
Process Node
Huaruizhipu Praxis One
TSMC 3nm (MediaTek Genio Pro 5100)
Traditional Robotic Edge Compute (e.g., NVIDIA Jetson / Discrete Setups)
Typically 4nm–8nm process nodes
NPU / AI Compute
Huaruizhipu Praxis One
106 TOPS (NPU 890 with Transformer Engine)
Traditional Robotic Edge Compute (e.g., NVIDIA Jetson / Discrete Setups)
Variable (e.g., 100–275 TOPS on Jetson Orin series)
CPU Core Config
Huaruizhipu Praxis One
8-core All-Big-Core (Cortex-X925 + X4 + A720, up to 260K DMIPS)
Traditional Robotic Edge Compute (e.g., NVIDIA Jetson / Discrete Setups)
Multi-core Arm Cortex-A78AE or x86 host processors
Vision / I/O Support
Huaruizhipu Praxis One
Up to 16 FHD camera virtual channels, 16-bit HDR, 8K video encode/decode
Traditional Robotic Edge Compute (e.g., NVIDIA Jetson / Discrete Setups)
8–16 camera interfaces via GMSL2 / MIPI CSI-2
Operating Temperature
Huaruizhipu Praxis One
Industrial-grade (−40°C to +105°C)
Traditional Robotic Edge Compute (e.g., NVIDIA Jetson / Discrete Setups)
Typically −25°C to +85°C (standard industrial SKUs)

Technical Deep Dive

  • SoC & Fabrication: Powered by the MediaTek Genio Pro 5100, manufactured using TSMC's 3nm fabrication technology.
  • NPU Architecture: MediaTek 8th-generation NPU 890 delivering up to 106 TOPS; includes a hardware-optimized Transformer engine capable of running 7B parameter LLMs (such as Qwen3 7B) at up to 23 tokens/second.
  • CPU Subsystem: 8-core Arm v9.2 configuration comprising 1× Cortex-X925, 3× Cortex-X4, and 4× Cortex-A720; features 12MB L3 cache and 10MB system-level cache, delivering up to 260K DMIPS compute for kinematics and path planning.
  • System Topology: Unified single-chip 'dual-brain' design integrating cognitive reasoning and real-time motion control loop execution to eliminate inter-chip bus latency.
  • Sensor & Interface Pipeline: Supports up to 16 FHD virtual camera input channels with native 16-bit hardware HDR; hardware acceleration for 8K video encoding and decoding; triple 4K display output.
  • Wireless & Operating Environment: Built-in Wi-Fi 7 and Bluetooth 6.0 with high-accuracy channel sounding; certified for industrial operating ranges between −40°C and +105°C.

Future ImplicationsAI analysis grounded in cited sources

Single-SoC robotic platforms will compress hardware iteration cycles for humanoid OEMs.
By consolidating cognitive AI and deterministic motion control into a pre-integrated 3nm SoC platform, manufacturers can bypass discrete bus debugging and shorten time-to-market by 9 to 12 months.
ARM-based mobile edge architectures will actively challenge discrete GPU supremacy in physical AI.
Energy-efficient 3nm NPU designs delivering over 100 TOPS enable local execution of Vision-Language-Action models within constrained thermal envelopes in robotics.

Timeline

2026-09
Huaruizhipu unveils Praxis One and Praxis Nano edge AI platforms at the 9th China Robot Summit ecosystem event in Hangzhou
Read original article on Pandaily

Weekly AI Recap

Read this week's curated digest of top AI events →

PostLinkedIn

AI-curated news aggregator. All content rights belong to original publishers.
Original source: Pandaily ↗

This is a summary, not the original. Read the source, or get the weekly briefing.

The weekly digest

One email a week. Unsubscribe anytime.