Skip to main content

Mobilint Regulus (South Korea NPU, 2023)

Product Overview

Mobilint is a South Korean AI chip company, founded 2019, Seoul. Regulus is its first commercial AI NPU, 2023-Q2 released, TSMC 5nm, 32 TOPS INT8 compute, 12W TDP (one of the most power-efficient NPUs in industry). Paired with Mobilint ARIES SDK + PyTorch / TensorFlow / ONNX compatibility.

Strategic significance: Mobilint is one of South Korea's AI Startup Three Swordsmen (alongside FuriosaAI and Rebellions), customers: KT (Korea Telecom), SK Group, Samsung SDS, Naver. Mass production in 2024, targeting NVIDIA L4 inference chip.

Core Specs

ItemParameter
ArchitectureMobilint Aries NPU
ProcessTSMC 5nm
NPU Core Count8x Aries NPU core
Per Core4 MB SRAM + 4 TOPS INT8
On-chip SRAM32MB
LPDDR5X16GB / 32GB
Memory Bandwidth204 GB/s
INT832 TOPS
FP1616 TFLOPS
BF1616 TFLOPS
TDP12W (one of the lowest data center NPU TDPs in industry)
Efficiency2.67 TOPS/W
Form FactorM.2 / PCIe Gen4 x4 / OAM
InterconnectPCIe Gen4
Mass Production2023-Q2
Unit Price~$300-500

Aries NPU Architecture

DimensionMobilint AriesTraditional GPU
Execution Modelscalar + vector + matrixCUDA cores + Tensor cores
Parallelismoperator-levelthread-level
On-chip Memory4 MB SRAM / core (32 MB total)shared L2 + registers
Dataflowdataflow optimizedcache lines
Power12W70-700W
TargetEdge + entry data centerall scenarios

Single Aries NPU Core

Single Aries Core:
- 4 MB SRAM (dual-port)
- 4 TOPS INT8 matrix unit
- 32-bit scalar + 256-bit SIMD vector
- 1 GHz clock

Mobilint Regulus:
- 8 x Aries Core = 32 TOPS INT8
- 8 x 4 MB = 32 MB on-chip SRAM
- 1 GHz x 8 = 8 GHz equivalent

South Korea AI Three Swordsmen Comparison

DimensionMobilint RegulusFuriosaAI RNGDRebellions RBLN
Compute INT832 TOPS1,000 TOPS16 TOPS
TDP12W450W15-30W
Efficiency2.67 TOPS/W2.22 TOPS/W1 TOPS/W
Memory32GB LPDDR5X512GB HBM316GB LPDDR5X
Price~$400~$22K~$1K
TargetEdge + entry DCdata center inferenceEdge + DC
Mass Production2023-Q22024-Q32024-Q2
CustomersKT / Samsung / NaverKT / SK / G42KT / SK / Samsung

Mobilint positioning: Edge + entry data center (vs FuriosaAI high-end DC vs Rebellions all scenarios). TDP 12W is the lowest among the Three Swordsmen.

vs NVIDIA L4

MetricMobilint RegulusNVIDIA L4Difference
INT832 TOPS485 TOPSL4 15x
TDP12W72WRegulus 6x power savings
Efficiency2.67 TOPS/W6.7 TOPS/WL4 2.5x
Memory32GB LPDDR5X24GB GDDR6Regulus 1.3x
Bandwidth204 GB/s300 GB/sL4 1.5x
Price~$400~$2,500Regulus 6x cheaper
SoftwareARIES SDKCUDAL4 mature

Regulus advantage: TDP 12W (L4 17%) + price 16% L4 + South Korea sovereign AI. L4 advantage: compute 15x + CUDA ecosystem + multi-cloud deployment.

Vendor Information

ItemContent
CompanyMobilint
FounderHyung-jin Kim (CEO, former Samsung semiconductor)
Founded2019
HeadquartersSeoul, South Korea + Silicon Valley, USA
Funding$100M+ (Series A 2023-Q1 led by: Korea Development Bank + SK Hynix)
Valuation (2025)~$500M
2024 Revenue~$15M
Employees~80
FabTSMC 5nm
Key CustomersKT (Korea Telecom), SK Group, Samsung SDS, Naver
Government SupportSouth Korea National AI Semiconductor Strategy
Statusmass production + 2026 planned IPO

Software Stack Mobilint ARIES

LayerToolDescription
AI frameworkMobilint ARIES SDKunified programming
PyTorch (ARIES backend)auto NPU mapping
TensorFlow Litecompatible
ONNX Runtimecompatible
CompilerARIES Compilermodel -> NPU
RuntimeARIES Runtimemulti-core scheduling
QuantizationARIES QuantINT8 automatic
Model ZooARIES ModelZoo50+ inference models

ARIES advantage: PyTorch native, compile 1-5 minutes (vs Cambricon NeuWare 1-2 hours), simple model deployment.

Use Cases

  • Edge AI inference (12W air-cooled)
  • South Korea sovereign AI (KT / SK / Naver customers)
  • Small model training (7B LLM quantized)
  • Entry data center inference (8-card cluster 256 TOPS)
  • Multimodal AI (image + text)
  • ADAS / autonomous driving (12W air-cooled)
  • AI training (inference + small training only)
  • CUDA proprietary workloads (requires ARIES porting)
  • International market (Korea / Asia primary)

Mobilint Product Line

ProductReleasedComputeTDPTarget
Aries (Regulus)2023-Q232 INT8 TOPS12WEdge + entry DC
Aries Pro (est.)2025 H264 INT8 TOPS20Whigh-end Edge
Aries 200 (est.)2026200 INT8 TOPS80Wdata center
Aries 1000 (est.)20271000 INT8 TOPS200Whigh-end data center

Mobilint vs Apple A18 Pro

MetricMobilint RegulusApple A18 Pro
INT832 TOPS35 TOPS
TDP12W8-10W
Efficiency2.67 TOPS/W3.5 TOPS/W
Memory32GB LPDDR5X8GB LPDDR5X
Price~$400iPhone $999+
Targetdata center + edgesmartphone

Mobilint advantage: 32GB memory (A18 Pro 8GB 4x) + data center deployment + low price. A18 Pro advantage: iPhone integration + 35 TOPS + 30% better efficiency.

South Korea AI Startup Three Swordsmen Strategic Significance

  • FuriosaAI: high-end data center inference (512GB HBM3)
  • Rebellions: edge + data center (16-400 TOPS evolution)
  • Mobilint: entry + edge (32 TOPS, 12W)

Three Swordsmen complementary: covering all South Korea AI scenarios (high-end / mid-range / entry / edge), South Korea national AI compute fully autonomous.

Key Features

  • 32 TOPS INT8: entry data center NPU
  • TDP 12W: lowest data center NPU TDP in industry
  • Price $400: cheapest in industry
  • South Korea + TSMC 5nm: mature manufacturing
  • ARIES SDK: PyTorch compatible
  • Drawbacks: low compute (vs L4 15x), 2-year ecosystem