Rebellions REBEL (Korean AI Chip・2024年)（リベリオンズリベル）
2024年Q2 Rebellions (韓国Seongnam・2020年Park Sungyun (元IBM Research / Intel Habana Labs) 創業・累計funding $400M+・Saudi Aramco Energy Ventures + Samsung投資・韓国Hyperscale AI Chip startup台頭最右翼) 発表のREBEL・1024 TOPS INT8 (Hailo-10H 40 TOPS の 25倍) AI推論Chip・Samsung Foundry 4nm process + Multi-Die CoWoS package + HBM3e 144GB onboard + 200W TDP + PCIe Gen 5 x16・$5,000-$8,000・NVIDIA H100直接競合のEnterprise Server-grade AI Chip・Korea Ministry of Science支援 + Samsung Foundry先進製造のNational AI Chip championship。
概要
Rebellions REBEL は、2024年Q2に韓国Seongnam拠点のRebellions Inc. (2020年Park Sungyun (CEO・元IBM Research AI / Intel Habana Labs Senior Engineer) + Oh Jin-hwook (CTO・元Samsung Research) 共同創業・累計funding $400M+ (Saudi Aramco Energy Ventures Lead Series B $124M + Samsung投資 + Korea Development Bank等・累計400 million USD scale)・韓国Hyperscale AI Chip startup台頭最右翼) が発表したEnterprise Server-grade AI推論Chip第2世代。前世代ATOM (2022年・Edge AI tier・40 TOPS INT8) の Hyperscale Server-grade Pivot後継。Samsung Foundry 4nm LPP (Low Power Plus) process + Multi-Die CoWoS-S package (TSMC技術相当・Samsung独自Multi-Die packaging) + HBM3e 144GB onboard memory (NVIDIA H200 141GB相当・H100 80GB の 1.8倍) + 200W TDP + PCIe Gen 5 x16 host接続・1024 TOPS INT8 + 512 TFLOPS BF16 (Hopper / Blackwell Tensor Core 競合 specification level) ・$5,000-$8,000/chip (NVIDIA H100 PCIe $25k-$30k の 1/4-1/5価格・Cost-down路線) ・NVIDIA H100直接競合のEnterprise Server-grade AI Inference Chip。Korea Ministry of Science 政府支援 + Samsung Foundry 4nm先進製造 + Saudi Aramco Energy Ventures Strategic Investor + Hyperscaler Customer (LG AI Research / SK Telecom / KT等の Korean Hyperscaler) Adoption実績の National AI Chip Champion位置づけ。
主な特徴・仕組み
- Performance: 1024 TOPS INT8 + 512 TFLOPS BF16 (NVIDIA H100 BF16 990 TFLOPS vs 約 0.52倍だが INT8 wise優位)
- Architecture: Rebellions独自NPU Architecture (32 RBP cores・各32 TFLOPS BF16 + 64 TOPS INT8)
- Process: Samsung Foundry 4nm LPP (Low Power Plus・前世代5nm vs Power efficiency改善)
- Packaging: Multi-Die CoWoS-S package (TSMC技術相当・Samsung独自Multi-Die packaging)
- Memory: HBM3e 144GB onboard (NVIDIA H200 141GB相当・H100 80GB の 1.8倍 + Bandwidth 4.8TB/s)
- Power: 200W TDP (NVIDIA H100 SXM 700W vs 1/3.5 + PCIe 350W vs 1/1.75)
- Form Factor: PCIe Gen 5 x16 Card + OAM (Open Compute Project Accelerator Module) variant
- Supported Models: GPT-4 class LLM Training + Inference + LLaMA 70B Q8 + Mistral 8x7B + Custom Customer models
- Software: Rebellions SDK (PyTorch / ONNX / TensorRT-LLM compatible) + Triton Server integration
- Inference Performance: GPT-4 class Inference ~50 tok/s (NVIDIA H100 ~100 tok/s vs 0.5倍・H100 1/4-1/5価格で2倍性能 = Cost-perfomance優位)
- Use Cases: LLM Training + Inference / Korean Hyperscaler / Saudi Government / Asia AI infrastructure
- Price: $5,000-$8,000/chip (NVIDIA H100 PCIe $25k-$30k vs 1/4-1/5)
スペック比較表
| Server-grade AI Chip (2024) | INT8 TOPS | Memory | Power | Process | Price |
|---|---|---|---|---|---|
| Rebellions REBEL | 1024 TOPS | HBM3e 144GB | 200W | Samsung 4nm | $5,000-$8,000 |
| NVIDIA H100 SXM | 1979 TOPS | HBM3 80GB | 700W | TSMC 4N | $25,000-$40,000 |
| NVIDIA H100 PCIe | 1513 TOPS | HBM3 80GB | 350W | TSMC 4N | $25,000-$30,000 |
| AMD MI300X | 1300 TOPS | HBM3 192GB | 750W | TSMC 5nm | $15,000-$20,000 |
| Intel Gaudi 3 | 1835 TFLOPS BF16 | HBM2e 128GB | 600W | TSMC 5nm | $15,000-$18,000 |
具体例・対応製品
- Rebellions REBEL PCIe Card: $5,000-$6,000・PCIe Gen 5 x16 + HBM3e 144GB
- Rebellions REBEL OAM Module: $6,000-$8,000・OAM 2.0 form factor + Higher density
- Rebellions SDK + Triton Integration: 無料・PyTorch / ONNX / TensorRT-LLM compat
- LG AI Research deployment: EXAONE 3.5 LLM Inference基盤 (2024年Q4本格稼働)
- 競合 NVIDIA H100 PCIe: $25k-$30k・1513 TOPS INT8 + HBM3 80GB + 350W (Single chip性能High但5倍高価)
自作PCでの選び方・注意点
Rebellions REBEL は「Korean Hyperscaler / Asian Cloud Provider (LG AI Research / SK Telecom / KT / Saudi Aramco) AI Infrastructure」「NVIDIA H100 1/4-1/5価格でCost-down LLM Inference」「Korea Government AI Sovereignty (NVIDIA dependency回避 Strategy)」用途のEnterprise Server-grade AI Chip。Consumer / 自作PC範囲外・カード単価 $5,000-$8,000で個人購入対象外・Korean Hyperscaler / Cloud Provider / Saudi National Cloud等 Tier-1 customer専用。NVIDIA H100 PCIe ($25k-$30k) と比較で Single chip性能 0.5-0.7倍 (1024 TOPS vs 1513 TOPS / 50 tok/s vs 100 tok/s GPT-4 class) だが価格 1/4-1/5 + HBM3e 144GB (H100 80GB の 1.8倍 = larger LLM model fits) + Power 200W (H100 PCIe 350W の 0.57倍) でCost-Performance + Energy efficiencyで国家strategic Investment価値あり。Korean Government AI Sovereignty Strategy (NVIDIA dependency回避・Korean ChampionとしてNational AI Chip育成) + Samsung Foundry 4nm先進製造のNational Manufacturing Sovereignty + Saudi Aramco Energy Ventures Strategic Partnership (Saudi Vision 2030 AI Sovereignty Strategy align) でAsia + Middle East AI Infrastructure市場で台頭。Korean Hyperscaler (LG AI Research EXAONE 3.5 LLM・2024年Q4 deploy・LG Group内全AI workload Rebellions REBEL基盤化方針) + SK Telecom / KT/Naver等の Korean AI Cloud台頭 + Asian non-NVIDIA AI infrastructure choices広がる。
関連用語との違い
- vs NVIDIA H100 PCIe: REBELは$5k-$8k + HBM3e 144GB + Korea製・H100は$25k-$30k + HBM3 80GB + NVIDIA Ecosystem深
- vs AMD MI300X: REBELは$5k-$8k + 200W Power efficient・MI300Xは$15k-$20k + HBM3 192GB Large + 750W
- vs Intel Gaudi 3: REBELはKorean製 + $5k-$8k Mid-range・Gaudi 3はIntel製 + $15k-$18k Mid-Premium
よくある質問（FAQ）
Q1: NVIDIA H100 5倍価格安でなぜ普及しない？ A: 製造Capacity制約 (Samsung Foundry 4nm wafer allocation限定 + Multi-Die CoWoS package supply constraint) + Software Ecosystem成熟度 (NVIDIA CUDA / TensorRT-LLM / NeMo / TensorRT-LLM比較Rebellions SDK 2024年Q2発売直後) + Customer Adoption time (Hyperscaler内workload migration 6-12ヶ月 timeline)。2025-2026年でCorrection期待。
Q2: Saudi Aramco Investmentの意味は？ A: Saudi Vision 2030 AI Sovereignty Strategy align ($30B+ AI Sovereignty fund) で Rebellions $124M Series B (2024年Q1) Lead Investment + Saudi National AI Cloud Provider deployment実績作り。Korean Champion brand → Asia + Middle East AI infrastructure expansion基盤。
Q3: Samsung Foundry 4nm の意義は？ A: TSMC 4N (NVIDIA H100 process) 相当先進製造でPerformance / Power efficiency NVIDIA H100 wise相当達成可能 + Korea National Manufacturing Sovereignty (NVIDIA = TSMC dependency回避・Korean Champion = Samsung Foundry成立で Korean Tech Stack完成)。
まとめ
Rebellions REBEL = 2024年Q2 Rebellions発表のKorean AI Chip。1024 TOPS INT8 + Samsung Foundry 4nm + Multi-Die CoWoS-S + HBM3e 144GB + 200W TDP + PCIe Gen 5 x16 + $5,000-$8,000・NVIDIA H100直接競合のEnterprise Server-grade AI Chip + Korean Government AI Sovereignty Strategy + Saudi Aramco Investment Asia National AI Champion位置づけ。