Skip to main content

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)

ItemParameter
ArchitectureRebellions RDU (Reconfigurable Dataflow)
ProcessTSMC 5nm
RDU Core Count8x RDU cores (32 MB SRAM per core)
On-chip SRAM256MB
LPDDR5X16GB
Memory Bandwidth102 GB/s
INT816 TOPS
BF168 TFLOPS
TDP15-30W (air-cooled, Edge-friendly)
Form FactorM.2 / PCIe Gen4 / OAM
InterconnectPCIe Gen4
Mass Production2024-Q2
Unit Price~$500-1500

ATOM Next-Gen (2025 Est.)

ItemParameter
ArchitectureRebellions RDU 2.0
ProcessTSMC 5nm
RDU Core Count32x RDU cores
HBM128GB HBM3
Memory Bandwidth~3 TB/s
BF16400 TFLOPS (competing with H100 inference)
INT8800 TOPS
TDP~250W
Mass Production2025-Q3 (estimated)

RDU Architecture

DimensionTraditional GPURebellions RDU
Execution ModelImperativeDataflow (Reconfigurable)
ParallelismThread-levelOperator-level (dataflow)
On-chip MemoryShared L2 + registersLarge distributed SRAM (32MB per core)
DataflowCache linesGraph streaming (reconfigurable)
Power70-700W15-30W (Edge-friendly)
Targetdata centeredge + data center
ReconfigurableCUDA programsdataflow 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

DimensionRebellions RBLN / ATOMFuriosaAI RNGD
Target Marketedge + data centerdata center inference
ArchitectureRDU (Reconfigurable Dataflow)TCP (Tensor Contraction)
Process5nm5nm
RBLN Compute16 INT8 TOPS-
ATOM Compute400 BF16 TF (est.)512 BF16 TF
ATOM Memory128GB HBM3512GB HBM3
TDP15-30W (RBLN) / 250W (ATOM)450W
DeploymentKT edge + cloudKT / SK / G42 data center
Funding$200M+$300M+
Valuation$1B+$1.5B+
IPO20262026-2027

Vendor Information

ItemContent
CompanyRebellions Inc.
Co-FounderSungkyung Kim (CEO, former Samsung semiconductor)
Co-FounderWoosung Kim (CTO)
Founded2020
HeadquartersSeoul, 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
FabTSMC 5nm
Key CustomersKT (Korea Telecom), SK Group, Samsung SDS, Naver HyperCLOVA X, Korea Investment & Securities
Government SupportSouth Korea National AI Semiconductor Strategy, K-Cloud project
Statuspreparing 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

ProductReleasedComputeTDPTarget
ISM (1st gen)20228 INT8 TOPS10WEdge early
RBLN (2nd gen)2024-Q216 INT8 TOPS15-30WEdge + data center
ATOM (3rd gen)2025-Q3 est.400 BF16 TF250Wdata center
ATOM Pro (est.)2026800 BF16 TF400Whigh-end data center

Rebellions vs NVIDIA L4 vs FuriosaAI

MetricRebellions RBLNNVIDIA L4FuriosaAI RNGD
INT816 TOPS485 TOPS1000 TOPS
TDP15-30W72W450W
Efficiency0.5-1 TOPS/W6.7 TOPS/W2.2 TOPS/W
Memory16GB LPDDR5X24GB GDDR6512GB HBM3
Price~$1K~$2.5K~$22K
TargetEdge / Inferencedata center inferencelarge 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).