Products | MLSoc Family

Real-time Performance, Ultra-low Power, and Seamless Deployment

Meet the industry’s first MLSoC family purpose-built for physical AI.

Gen 1

MLSoC

The first generation SoC is ideal for Vision ML at the edge (robotics, drones, healthcare, etc.)
Gen 2

MLSoC Modalix

The second generation SoC is ideal for Multimodal ML and GenAI in edge/cloud systems across broader verticals
Watch Video (2:15)
Join SoM EAP
GEN 1
GEN 2
MLSoC
MLSoC Modalix
M50 / M100 / M200
TOPS
50
50/100/200
Machine Learning Accelerator (MLA)
MLA optimized for CNN-based Models. INT8 and INT16 support.
MLA enhanced with hardware blocks to accelerate GenAI computations. Supports mixed precision: Floating point (BFLOAT16) and INT8 and INT16
Application Processor
Arm A65 x 4 dual-threaded 1.15GHz; 15K DMIPS
Arm A65 x 8 dual threaded 1.4GHz; 32K DMIPS
MIPI Support
N/A
4 independent CSI 2 interfaces each with 4x 2.5Gbps lanes
ISP
N/A
Arm C-71 1.2GHz, supporting 8-24 bit RAW including RGBIR
Vision Processor
Synopsys ARC EV-74 800 MHz
Synopsys ARC EV-74 1GHz
Video Decoder
H.264/265 4KP60 x 1
H.264/265/AV1 4KP60 x 1
Video Encoder
H.264 4KP30 x 1
H.264 4KP30 x 1
PCIe
Gen 4 x 8 lanes endpoint only
Gen 5 x 8 lanes Root Complex and Endpoint
Ethernet
1GbE x 4
10GbE x 4
On-chip Memory
4 MB
8 MB
DRAM
32-bit LPDDR4 x 4
128 bit (8 channels) 32 & 64 bit LPDDR4/4X/5
Manufacturing Process Node
TSMC 16nm
TSMC 6nm
Model Types
CNNs, Transformers
CNNs, Transformers, LLMs, LMMs & GenAI
TOPS
Machine Learning Accelerator (MLA)
Application Processor
MIPI Support
ISP
Vision Processor
Video Decoder
Video Encoder
PCIe
Ethernet
On-chip Memory
DRAM
Manufacturing Process Node
Model Types
50
MLA optimized for CNN-based Models. INT8 and INT16 support.
Arm A65 x 4 dual-threaded 1.15GHz; 15K DMIPS
N/A
N/A
Synopsys ARC EV-74 800 MHz
H.264/265 4KP60 x 1
H.264 4KP30 x 1
Gen 4 x 8 lanes endpoint only
1GbE x 4
4 MB
32-bit LPDDR4 x 4
TSMC 16nm
CNNs, Transformers
M50
M100
M200
50
100
200
MLA enhanced with hardware blocks to accelerate GenAI computations. Supports mixed precision: Floating point (BFLOAT16) and INT8 and INT16
Arm A65 x 8 dual threaded 1.4GHz; 32K DMIPS
4 independent CSI 2 interfaces each with 4x 2.5Gbps lanes
Arm C-71 1.2GHz, supporting 8-24 bit RAW including RGBIR
Synopsys ARC EV-74 1GHz
H.264/265/AV1 4KP60 x 1
H.264 4KP30 x 1
Gen 5 x 8 lanes Root Complex and Endpoint
10GbE x 4
8 MB
128 bit (8 channels) 32 & 64 bit LPDDR4/4X/5
TSMC 6nm
CNNs, Transformers, LLMs, LMMs & GenAI