top of page
pic_001.png

Modula-X01™AR/VR Edge Graphics Accelerator

AR/VR goes Real-Time at the Edge Multi-user, venue-scale experiences powered by a 16-tile edge accelerator.

​

Modula-X01 is a 16-tile graphics and vision accelerator for AR/VR edge computers; the firmware/runtime presents one unified pool so teams deploy without touching low-level settings.

截圖 2025-11-07 下午3.50.04.jpg
pic_001.png

Market Trend & Needs of Interactive AR/VR via Edge Computer

Macro Trends

​

  • Latency budgets push compute off cloud, off headset, onto the “near edge.”

Motion/interaction loops need sub-20 ms end-to-end; placing compute in-room/onsite removes WAN jitter without burdening the HMD.

​

  • Multi-sensor environments are becoming standard.

Edge nodes ingest multi-camera video, depth/LiDAR, IMU, mic arrays—doing fusion, tracking, and mapping once for multiple users.

​

  • Advanced 3D techniques are moving real-time

Real-time reconstruction, upscaling, reprojection/compositing, super-res, depth fusion, SLAM and scene reconstruction require steady throughput and predictable latency.

​

  • Enterprises want privacy + reliability

Content and biometrics stay onsite; sessions must run offline if the WAN hiccups.

Demand Markets

​

  • Venues

retail mirrors, showrooms, museums & exhibits; Online/Location-Based Gaming lounges and esports arenas; industrial guided-assembly cells; education/medical simulation labs.

​

  • Industrial

guided assembly, remote expert, safety training cells.

​

  • Healthcare & education

3D anatomy visualization, simulation labs.

​

  • Representative workloads:

hand/pose/eye tracking, VIO/SLAM, depth fusion/meshing, real-time reconstruction, upscaling, reprojection/compositing, depth fusion/meshing, tracking/SLAM, spatial audio, denoise/super-resolution, spatial audio; plus low-latency player tracking & scene compositing for online/arcade gaming, and multi-stream encode/stream to HMDs/displays.

Key Features

Purpose-built for venue-scale AR/VR: stable < 20 ms interaction, 1–4 users per box, on-prem privacy.

l16 compute tiles, one pool (FW/SDK abstracts everything)

lDeterministic frame timing for 60–120 Hz AR/VR

lSpatial compute engines for sensor fusion, VIO, 3D geometry & compositing

lImage/Video pipeline: denoise, HDR, color, warp, low-variance encode/decode

lOn-chip SRAM + DDR4/DDR5 (ECC optional) for predictable latency

lMulti-camera readiness (MIPI-CSI; optional USB/Ethernet ingress)

lPCIe Gen4 x8 host (CXL-capable features under vendor firmware; not user-programmable)

lQuiet edge-box thermals with stable power profile

lUnified SDK:  pipeline import, operator packs, session templates (Solo/Dual/Quad)

lSecure boot & signed updates (no user crypto acceleration)

Technical Needs and Solutions
AR/VR Edge Computer technical needs >> Modula-X01 Solves It

 

Deterministic low latency (stable ms/frame for hand/eye/pose loops)

Dataflow scheduler + on-chip SRAM tiling; bounded queueing; predictable kernel times under load.

Consistent 60–120 Hz pipelines with tight p95/p99

​

Fixed-latency operator library (tracking/SLAM/reprojection; priority streams and budgeted DMA)

​

Multi-stream ingest & fusion (4–12 camera/depth feeds at HD/FHD)

Parallel DMA engines, pre/post pipelines, zero-copy buffer views; fused ops for resize/denoise/pack

​

Tight thermal/acoustic limits (~30–100 W node)

High perf/Watt PE arrays, fine-grain power/clock gating, smooth power profile (avoids fan ramp spikes)

​

 

Scalability to 1–4 HMDs/displays with QoS

Multi-queue execution with per-session quotas; throughput isolation for concurrent users

​

Robust VIO/SLAM & real-time 3D workloads

Optimized kernels for VIO/SLAM, depth fusion/meshing; classical 3D & vision (FP16/FP32 as applicable)

​

Low-variance encode/stream to HMDs

Overlapped compute–DMA–encode scheduling; large coalesced transfers; PCIe Gen4 x8 bandwidth headroom

On-prem privacy & offline resiliency

All processing local; vendor-controlled secure boot/update; no user-exposed general crypto API

​

DevOps-friendly tooling (fast time-to-demo)

Linux SDK, Graph import, sample graphs, timeline profiler; containerized deployment & remote update

High-resolution photo of a precision ASIC ic on a PCB with real lighting and lightning in

Target Applications 

機械工程

Engineering / Industrial (CAE/Simulation)

• Finite Element Method (FEM) → stress/strain analysis, structural engineering

• Computational Fluid Dynamics (CFD) → airflow, aerodynamics, thermal simulations

• Electromagnetic Simulation → antenna design, EMC/EMI testing

• Computer-Aided Design (CAD) optimizations → parametric modeling, mesh refinement

• Digital Twin training → real-time system replicas of factories, machines, or cities

Image by Ritu Chauhan

Medical / Healthcare

• 3D Medical Imaging Reconstruction → CT/MRI → volumetric training models

• Disease Classification / Detection → training models on medical images (X-ray, pathology slides)

• Protein Structure / Drug Modeling → molecular dynamics, docking simulations

• Biomechanics Simulation → bone, joint, and tissue stress models

Image by BoliviaInteligente

Scientific Computing

• Physics Simulations → quantum chemistry, plasma, particle interactions

• Materials Science → nanomaterial property modeling, semiconductor device modeling

• Climate & Environmental Modeling → weather, pollution dispersion, geological surveys

Image by Google DeepMind

3D Graphics / Computer Vision 

•   3D Object Recognition → training models on point clouds, CAD datasets

•   Point Cloud Transformers (LIDAR/SLAM) → robotics, autonomous driving

•   Neural Radiance Fields (NeRF) → training for 3D scene reconstruction

•   Generative 3D Models (Diffusion → Mesh) → asset generation for AR/VR/XR

•   3D Morphology Analysis → design optimization, topological learning

High-resolution photo of a precision ASIC ic on a PCB with real lighting and lightning in

Performance & Determinism

Stable interactive pipelines for AR/VR edge — frame-time consistency over peak TOPS

Pink Poppy Flowers

Reference System & Integration

截圖 2025-11-07 下午3.32.09.jpg

Deployment Templates (Multi-Instance)

Stable interactive pipelines for AR/VR edge — frame-time consistency over peak TOPS

截圖 2025-11-07 下午3.36.39.jpg
Image by Vishnu Mohanan

Roadmap - MODULA Series

Modula-X01 16 Tiles 

(validation & pilot)

  • Process: N16 • Tiles: 16 (unified pool)

  • Interfaces: PCIe Gen4 x8, MIPI‑CSI multi‑cam, encode/stream

  • Profiles: Quiet ~80 W (steady), Boost ≤110 W (short)

  • Latency: < 18 ms motion‑to‑photon (target)

  • Frame budgets: 90/120 Hz p95 & p99 under target

  • Users per box: up to 4 (Quad Balanced template)

  • Status: Validation vehicle; pilot deployments

Modula-X01L 8 Tiles

(compact / cost‑optimized)

  • Process: N16 • Tiles: 8 (same SDK)

  • Interfaces: PCIe Gen4 x8, MIPI‑CSI multi‑cam

  • Profiles: Quiet 55–65 W (steady), Boost ≤80 W (short)

  • Latency/Frame budgets: same as X01 at scaled loads

  • Users per box: up to 2 (Dual Pro template)

  • Fit: portable carts, kiosks, cost‑sensitive rollouts

  • Status: Design Ready, sample target: 2027H1

Modula-X02 16 Tiles

(Performance lift)

  • Process: N6 • Tiles: 16 (software‑compatible)

  • Perf/Watt: +40–60% vs X01 at same power (target)

  • Interfaces: PCIe Gen5 x8 (backward‑compatible), higher MIPI bandwidth

  • Users per box: up to 6 (subject to workload)

  • Ingest: up to 8× RGB + 4× Depth (HD) sustained (target)

  • Profiles: Quiet ~85–95 W, Boost ≤130 W (guardrailed)

  • Status: Architecture in progress; samples target 2028H1

Architecture

Modula-X01
AR/VR Edge 3D Graphics Accelerator

Pink Poppy Flowers
bottom of page