AISiMa.ai Modalix

Modalix MLSoC

MLSoC

[SiMa.ai]

Modalix is SiMa.ai's second-generation machine-learning system-on-chip: 50 TOPS at under 10 W in a 25 x 25 mm package, with Arm application cores, a computer vision unit, an image signal processor and 4K video codecs on one chip, for boards designed around the chip itself.

System Overview
  • Form Factor

    FCBGA chip, 25 x 25 mm

  • AI Performance

    50 TOPS

  • Power

    Under 10 W

  • Processor

    8x Arm Cortex-A65

Product Information
  • Model

    Modalix MLSoC

  • Vendor

    SiMa.ai

  • Category

    On-Premise AI

  • Product Group

    SiMa.ai Modalix

  • Type

    MLSoC

AceTec Support
  • Consulting

    Requirements analysis and custom system design

  • Implementation

    Integration guidance and lifecycle technical support

  • Verification

    Testing against specifications and industry standards

Features

  • Second-generation SiMa.ai MLSoC
  • 50 TOPS (BF16, INT8, INT16) at under 10 W
  • 8x Arm Cortex-A65 @ 1.4 GHz (octal core, 16 threads)
  • Quad-core Synopsys ARC EV74 computer vision unit, up to 720 16-bit GOPS
  • Arm Mali-C71AE image signal processor @ 1.2 GHz, 24-bit wide dynamic range
  • 4K60 video decode (H.264, H.265, AV1, MJPEG) and encode (H.264, H.265, MJPEG)
  • x8 PCIe Gen5 root complex and endpoint with bifurcation; 4x 10G Ethernet; 4x MIPI CSI-2 (4 lanes each)
  • 128-bit LPDDR5/4x/4 up to 6,400 Mbps; 8 MB on-chip memory
  • Hardware secure boot with OTP secure keys; TrustZone
  • 25 x 25 mm FCBGA (1369 balls); commercial (0 to +70 °C) and industrial (-40 to +85 °C)

Overview

One chip for Physical AI

Modalix combines a machine-learning accelerator, an Arm application processor, a DSP-based computer vision unit, an image signal processor and video encode and decode engines on one chip, joined by an internal, secure network-on-chip. It runs multimodal transformers, LLMs, LMMs and generative AI, as well as CNNs and traditional computer vision, without cloud connectivity.

Chip-down design

For a board built around the chip itself, SiMa.ai provides the chip-down design guide, SoC datasheet and design packages under NDA. The same chip is on the Modalix SoM, in the DevKit and on the PCIe card.

Software

SiMa.ai's Palette software suite, with the Palette Neat agentic development environment, programs the application processor, computer vision unit and ML accelerator together in end-to-end pipelines: bring an ONNX model, tune it with SiMa.ai's Model Compiler and build the application with frameworks such as OpenCV.

Specifications

Processing

Item

Content

ML Accelerator

50 TOPS (BF16, INT8, INT16); LLMs, neural networks and GenAI

Application Processor

8x Arm Cortex-A65 @ 1.4 GHz, 32k Dhrystone MIPS

Computer Vision Unit

4-core Synopsys EV74 DSP @ 1 GHz, 720 16-bit GOPS

Image Signal Processor

Arm Mali-C71AE @ 1.2 GHz, low latency

Video Decode

H.264 / H.265 / AV1, 4K60

Video Encode

H.264 / H.265, 4K60

On-Chip Memory

8 MB

Memory and Interfaces

Item

Content

Memory

4x LPDDR5 interfaces; 128-bit LPDDR5/4x/4 (4x 32-bit or 2x 64-bit), up to 6,400 Mbps

PCIe

x8 PCIe Gen5, root complex and endpoint, with bifurcation

Ethernet

4x 10G

Camera

4x MIPI CSI-2, 4 lanes each

Storage

SDIO / eMMC; 2x Octal SPI

Configurable I/O

64: up to 64 GPIO, 32 UART, 16 SPI, 32 I2C

Security

Item

Content

Boot and Security

Hardware secure boot engine: chain of trust, the boot image decrypted and authenticated with OTP secure keys

TrustZone, encryption, eKeys

Package

Item

Content

Package

25 x 25 mm FCBGA, 1369 balls

Commercial

0 to +70 °C — SM2-201B-C0AA0-000

Industrial

-40 to +85 °C — SM2-201B-I0AA0-000

Design Package

Chip-down design guide, SoC datasheet and design packages from SiMa.ai under NDA

Resources