Rebellions RBLN / ATOM (South Korea AI Inference, 2024)
Product Overview
Rebellions is a South Korean AI inference chip company, founded 2020, Seoul. RBLN (Rebellions Neural Processing Unit) is its data center AI inference chip, 2024-Q2 released, TSMC 5nm, 16 INT8 TOPS (estimated), 15-30W TDP. Paired with RDU (Reconfigurable Dataflow Unit) heterogeneous dataflow architecture + SDK compatible with PyTorch / ONNX.
Strategic significance: Rebellions is one of South Korea's AI Startup Duo (alongside FuriosaAI), first customers: KT (Korea Telecom), SK Group, Samsung, Naver. ATOM is the 2025 next-gen chip (128GB HBM, 400 TFLOPS), competing with NVIDIA H100 / L4.
Core Specs (RBLN Current)
| Item | Parameter |
|---|---|
| Architecture | Rebellions RDU (Reconfigurable Dataflow) |
| Process | TSMC 5nm |
| RDU Core Count | 8x RDU cores (32 MB SRAM per core) |
| On-chip SRAM | 256MB |
| LPDDR5X | 16GB |
| Memory Bandwidth | 102 GB/s |
| INT8 | 16 TOPS |
| BF16 | 8 TFLOPS |
| TDP | 15-30W (air-cooled, Edge-friendly) |
| Form Factor | M.2 / PCIe Gen4 / OAM |
| Interconnect | PCIe Gen4 |
| Mass Production | 2024-Q2 |
| Unit Price | ~$500-1500 |
ATOM Next-Gen (2025 Est.)
| Item | Parameter |
|---|---|
| Architecture | Rebellions RDU 2.0 |
| Process | TSMC 5nm |
| RDU Core Count | 32x RDU cores |
| HBM | 128GB HBM3 |
| Memory Bandwidth | ~3 TB/s |
| BF16 | 400 TFLOPS (competing with H100 inference) |
| INT8 | 800 TOPS |
| TDP | ~250W |
| Mass Production | 2025-Q3 (estimated) |
RDU Architecture
| Dimension | Traditional GPU | Rebellions RDU |
|---|---|---|
| Execution Model | Imperative | Dataflow (Reconfigurable) |
| Parallelism | Thread-level | Operator-level (dataflow) |
| On-chip Memory | Shared L2 + registers | Large distributed SRAM (32MB per core) |
| Dataflow | Cache lines | Graph streaming (reconfigurable) |
| Power | 70-700W | 15-30W (Edge-friendly) |
| Target | data center | edge + data center |
| Reconfigurable | CUDA programs | dataflow graph reconfiguration |
RDU Core
Single RDU Core:
- 32 MB SRAM
- 256 8-bit MACs
- Dataflow scheduler
- Scalar + vector + tensor units
RBLN 8-core:
- 8 x 32 MB = 256 MB SRAM
- 8 x 256 MAC = 2048 MACs
- Dataflow interconnect
South Korea AI Startup Duo
| Dimension | Rebellions RBLN / ATOM | FuriosaAI RNGD |
|---|---|---|
| Target Market | edge + data center | data center inference |
| Architecture | RDU (Reconfigurable Dataflow) | TCP (Tensor Contraction) |
| Process | 5nm | 5nm |
| RBLN Compute | 16 INT8 TOPS | - |
| ATOM Compute | 400 BF16 TF (est.) | 512 BF16 TF |
| ATOM Memory | 128GB HBM3 | 512GB HBM3 |
| TDP | 15-30W (RBLN) / 250W (ATOM) | 450W |
| Deployment | KT edge + cloud | KT / SK / G42 data center |
| Funding | $200M+ | $300M+ |
| Valuation | $1B+ | $1.5B+ |
| IPO | 2026 | 2026-2027 |
Vendor Information
| Item | Content |
|---|---|
| Company | Rebellions Inc. |
| Co-Founder | Sungkyung Kim (CEO, former Samsung semiconductor) |
| Co-Founder | Woosung Kim (CTO) |
| Founded | 2020 |
| Headquarters | Seoul, South Korea + Silicon Valley, USA |
| Funding | $200M+ (Series B 2024-Q2 led by: KT + Korea Development Bank) |
| Valuation (2025) | $1B+ (unicorn) |
| 2024 Revenue | ~$25M |
| Employees | ~150 |
| Fab | TSMC 5nm |
| Key Customers | KT (Korea Telecom), SK Group, Samsung SDS, Naver HyperCLOVA X, Korea Investment & Securities |
| Government Support | South Korea National AI Semiconductor Strategy, K-Cloud project |
| Status | preparing 2026 IPO |
Use Cases
- ✅ Edge AI inference (15-30W air-cooled)
- ✅ South Korea sovereign AI (KT / SK / Naver customers)
- ✅ LLM inference (ATOM 128GB HBM fits 70B)
- ✅ Data center inference (ATOM 250W energy-efficient)
- ✅ Multimodal AI (image + text)
- ❌ AI training (inference only)
- ❌ CUDA proprietary workloads (requires SDK porting)
- ❌ International market (Korea / Asia primary)
Key Features
- RDU dataflow architecture: reconfigurable + Edge-friendly
- 15-30W TDP: one of the lowest data center inference TDP in industry
- ATOM 128GB HBM: 2025 competing with H100
- South Korea national AI: KT / SK / Naver customer base
- Drawbacks: compute below H100 (2x), 2-year SDK ecosystem
Rebellions Product Line
| Product | Released | Compute | TDP | Target |
|---|---|---|---|---|
| ISM (1st gen) | 2022 | 8 INT8 TOPS | 10W | Edge early |
| RBLN (2nd gen) | 2024-Q2 | 16 INT8 TOPS | 15-30W | Edge + data center |
| ATOM (3rd gen) | 2025-Q3 est. | 400 BF16 TF | 250W | data center |
| ATOM Pro (est.) | 2026 | 800 BF16 TF | 400W | high-end data center |
Rebellions vs NVIDIA L4 vs FuriosaAI
| Metric | Rebellions RBLN | NVIDIA L4 | FuriosaAI RNGD |
|---|---|---|---|
| INT8 | 16 TOPS | 485 TOPS | 1000 TOPS |
| TDP | 15-30W | 72W | 450W |
| Efficiency | 0.5-1 TOPS/W | 6.7 TOPS/W | 2.2 TOPS/W |
| Memory | 16GB LPDDR5X | 24GB GDDR6 | 512GB HBM3 |
| Price | ~$1K | ~$2.5K | ~$22K |
| Target | Edge / Inference | data center inference | large model inference |
Rebellions RBLN advantage: 15-30W TDP (L4 40% energy savings) + low price (L4 40% price). L4 advantage: 485 TOPS (RBLN 30x compute) + CUDA ecosystem. RNGD advantage: 512GB HBM3 (RBLN 32x memory).
Related Products
- FuriosaAI RNGD - South Korea duo
- Groq LPU v2 - LPU inference
- SambaNova SN40L - Dataflow
- Blaize Xplorer X1600 - Edge AI
- Hailo-8 / Hailo-15 - Edge AI
- Qualcomm AI 100 - Edge AI
- NVIDIA L4 - edge inference
- Cambricon MLU 590 - Chinese comparison