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
| Item | Parameter |
|---|
| Architecture | Mobilint Aries NPU |
| Process | TSMC 5nm |
| NPU Core Count | 8x Aries NPU core |
| Per Core | 4 MB SRAM + 4 TOPS INT8 |
| On-chip SRAM | 32MB |
| LPDDR5X | 16GB / 32GB |
| Memory Bandwidth | 204 GB/s |
| INT8 | 32 TOPS |
| FP16 | 16 TFLOPS |
| BF16 | 16 TFLOPS |
| TDP | 12W (one of the lowest data center NPU TDPs in industry) |
| Efficiency | 2.67 TOPS/W |
| Form Factor | M.2 / PCIe Gen4 x4 / OAM |
| Interconnect | PCIe Gen4 |
| Mass Production | 2023-Q2 |
| Unit Price | ~$300-500 |
Aries NPU Architecture
| Dimension | Mobilint Aries | Traditional GPU |
|---|
| Execution Model | scalar + vector + matrix | CUDA cores + Tensor cores |
| Parallelism | operator-level | thread-level |
| On-chip Memory | 4 MB SRAM / core (32 MB total) | shared L2 + registers |
| Dataflow | dataflow optimized | cache lines |
| Power | 12W | 70-700W |
| Target | Edge + entry data center | all 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
| Dimension | Mobilint Regulus | FuriosaAI RNGD | Rebellions RBLN |
|---|
| Compute INT8 | 32 TOPS | 1,000 TOPS | 16 TOPS |
| TDP | 12W | 450W | 15-30W |
| Efficiency | 2.67 TOPS/W | 2.22 TOPS/W | 1 TOPS/W |
| Memory | 32GB LPDDR5X | 512GB HBM3 | 16GB LPDDR5X |
| Price | ~$400 | ~$22K | ~$1K |
| Target | Edge + entry DC | data center inference | Edge + DC |
| Mass Production | 2023-Q2 | 2024-Q3 | 2024-Q2 |
| Customers | KT / Samsung / Naver | KT / SK / G42 | KT / 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
| Metric | Mobilint Regulus | NVIDIA L4 | Difference |
|---|
| INT8 | 32 TOPS | 485 TOPS | L4 15x |
| TDP | 12W | 72W | Regulus 6x power savings |
| Efficiency | 2.67 TOPS/W | 6.7 TOPS/W | L4 2.5x |
| Memory | 32GB LPDDR5X | 24GB GDDR6 | Regulus 1.3x |
| Bandwidth | 204 GB/s | 300 GB/s | L4 1.5x |
| Price | ~$400 | ~$2,500 | Regulus 6x cheaper |
| Software | ARIES SDK | CUDA | L4 mature |
Regulus advantage: TDP 12W (L4 17%) + price 16% L4 + South Korea sovereign AI.
L4 advantage: compute 15x + CUDA ecosystem + multi-cloud deployment.
| Item | Content |
|---|
| Company | Mobilint |
| Founder | Hyung-jin Kim (CEO, former Samsung semiconductor) |
| Founded | 2019 |
| Headquarters | Seoul, 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 |
| Fab | TSMC 5nm |
| Key Customers | KT (Korea Telecom), SK Group, Samsung SDS, Naver |
| Government Support | South Korea National AI Semiconductor Strategy |
| Status | mass production + 2026 planned IPO |
Software Stack Mobilint ARIES
| Layer | Tool | Description |
|---|
| AI framework | Mobilint ARIES SDK | unified programming |
| PyTorch (ARIES backend) | auto NPU mapping |
| TensorFlow Lite | compatible |
| ONNX Runtime | compatible |
| Compiler | ARIES Compiler | model -> NPU |
| Runtime | ARIES Runtime | multi-core scheduling |
| Quantization | ARIES Quant | INT8 automatic |
| Model Zoo | ARIES ModelZoo | 50+ 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
| Product | Released | Compute | TDP | Target |
|---|
| Aries (Regulus) | 2023-Q2 | 32 INT8 TOPS | 12W | Edge + entry DC |
| Aries Pro (est.) | 2025 H2 | 64 INT8 TOPS | 20W | high-end Edge |
| Aries 200 (est.) | 2026 | 200 INT8 TOPS | 80W | data center |
| Aries 1000 (est.) | 2027 | 1000 INT8 TOPS | 200W | high-end data center |
Mobilint vs Apple A18 Pro
| Metric | Mobilint Regulus | Apple A18 Pro |
|---|
| INT8 | 32 TOPS | 35 TOPS |
| TDP | 12W | 8-10W |
| Efficiency | 2.67 TOPS/W | 3.5 TOPS/W |
| Memory | 32GB LPDDR5X | 8GB LPDDR5X |
| Price | ~$400 | iPhone $999+ |
| Target | data center + edge | smartphone |
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