Axelera Metis AI (214 TOPS Edge AI Server Chip・2024年)（アクセレラメティス）
2024年Q2 Axelera AI (Netherlands Eindhoven・2021年Fabrizio Del Maffeo (元IBM Watson AI) 創業・Edge AI Server用途特化・累計funding $120M+) 発表のMetis AI・Edge AI Server向け214 TOPS INT8 (Hailo-10H 40 TOPS の 5倍以上) AI推論Chip・Digital In-Memory Computing (D-IMC) Architecture (独自Patent技術・SRAM内Computing) + LPDDR4x onboard memory + 7-10W TDP + M.2 + PCIe HHHL (Half-Height Half-Length card) Multi form factor + Vision AI / Generative AI / LLM対応・$400-$800・Edge AI Server / AI Workstation / Smart City Camera Infrastructure用途主軸。
概要
Axelera Metis AI は、2024年Q2にNetherlands Eindhoven拠点のAxelera AI (2021年Fabrizio Del Maffeo (CEO・元IBM Watson AI Lead) + Evangelos Eleftheriou (CTO・元IBM Research) 共同創業・累計funding $120M+) が発表したEdge AI Server向け高性能AI推論Chip。Hailo-10H (40 TOPS INT8 + Mainstream Edge AI Chip $200-$400) との差別化はEdge AI Server / AI Workstation Tier (Mid-range performance・Hailo Edge AI Mass-market segmentと別Tier上位)・214 TOPS INT8 (Hailo-10H 40 TOPS の 5倍以上・NVIDIA Jetson AGX Orin 275 TOPS と類似性能水準) ・Digital In-Memory Computing (D-IMC) Architecture (Axelera独自Patent技術・SRAM内Computing・Von Neumann bottleneck回避 + 10-30倍Energy efficiency improvement)・LPDDR4x 16GB onboard memory + 7-10W TDP + M.2 (Compact form factor) + PCIe HHHL (Half-Height Half-Length card・Server / Workstation insert可能) 両form factor提供 + Vision AI / Generative AI / LLM (Llama 3 8B Q4 / 13B Q8) 全対応・$400-$800/chip ($600 target volume pricing) ・Edge AI Server (Smart Factory / Smart City Camera Infrastructure / Edge AI Workstation) 主軸用途。
主な特徴・仕組み
- Performance: 214 TOPS INT8 (Hailo-10H 40 TOPS の 5倍以上)
- Architecture: Digital In-Memory Computing (D-IMC) ・SRAM内Computing・Von Neumann bottleneck回避
- Energy Efficiency: 10-30 TOPS/W (Hailo-10H 11 TOPS/W同等・Architecture理論value)
- Power: 7-10W TDP (Hailo-10H 3.5W vs 2-3倍だが性能5倍で総合効率優位)
- Memory: LPDDR4x 16GB (Hailo-10H DDR4 8GB vs 2倍容量)
- Form Factor: M.2 2280 (Compact) + PCIe HHHL (Half-Height Half-Length・Server insert)
- Supported Models: Llama 3 8B Q4 / Llama 3 13B Q8 / Llama 3.2 3B / Mistral 7B / SDXL / LLaVA 13B / Whisper Medium
- Inference Performance: Llama 3 8B Q4 token gen ~30-40 tok/s (Hailo-10H 15 tok/s vs 2-2.5倍高速)
- Generative AI: LLM / Diffusion / VLM all-modality対応
- Voyager SDK: Compilation toolchain・PyTorch / ONNX / TensorFlow → Axelera native format変換
- Voyager Model Zoo: Pre-compiled Llama / Mistral / SDXL等公開
- Driver: Linux Kernel driver / Windows driver (M.2 + PCIe HHHL两対応)
- Use Cases: Edge AI Server / Smart City Camera Infrastructure / AI Workstation / Industrial Vision AI
- Price: $400-$800/chip ($600 target volume pricing)
スペック比較表
| Edge AI Server Chip (2024) | Performance | Architecture | Power | Memory | Price |
|---|---|---|---|---|---|
| Axelera Metis | 214 TOPS INT8 | D-IMC SRAM内 | 7-10W | LPDDR4x 16GB | $400-$800 |
| Hailo-10H | 40 TOPS INT8 + 20 TFLOPS BF16 | Dataflow | 3.5W | DDR4 8GB | $200-$400 |
| NVIDIA Jetson AGX Orin 64GB | 275 TOPS INT8 (Sparsity 200) | Ampere GPU + NVDLA | 15-60W | LPDDR5 64GB | $1,999 |
| Tenstorrent Blackhole | 745 TOPS INT8 | RISC-V Many-core | 200W | GDDR6 32GB | $1,500-$3,000 |
| Groq LPU | 750 TFLOPS LLM Inference (Cloud) | TSP Tensor Streaming | 240W | SRAM内 | Cloud only |
具体例・対応製品
- Axelera Metis M.2 2280: $400-$600・PCIe Gen 4 x4 + LPDDR4x 16GB・Compact deployment
- Axelera Metis PCIe HHHL Card: $600-$800・PCIe Gen 4 x8 + LPDDR4x 16GB・Server / Workstation insert
- Axelera Voyager SDK: 無料・PyTorch / ONNX → Axelera native format conversion + Pre-compiled models
- Axelera Reference Designs: Smart City Camera Infrastructure / Industrial Vision AI applications
- 競合 Hailo-10H: $200-$400・40 TOPS Mainstream Edge AI・$200低価格帯
自作PCでの選び方・注意点
Axelera Metis AI は「Edge AI Server / AI Workstation deployment (Mid-range Performance 200 TOPS+)」「Smart City / Industrial Vision AI Infrastructure (24/7運用・Mass scale)」「Edge LLM 7-13B parameter Local Inference」用途の本命Edge AI Server Chip。Hailo-10H (40 TOPS Mainstream Edge AI $200-$400) の Edge AI Mass-market segment と異なり、Server-grade performance (214 TOPS = NVIDIA Jetson AGX Orin 275 TOPS類似) + LPDDR4x 16GB Memory (Llama 3 13B Q8 model size fits)で Edge AI Server / AI Workstation deployment向け Class。Digital In-Memory Computing (D-IMC) Architecture (Axelera独自Patent技術) はVon Neumann bottleneck (Memory <-> Compute data transfer ボトルネック) を SRAM内Computing で根本解決 + 10-30 TOPS/W Energy efficiency実現・Edge AI Server 24/7運用 でPower budget厳しい Smart City Camera Infrastructure / Industrial Vision AI用途に適合。一方NVIDIA Jetson AGX Orin 64GB ($1,999・275 TOPS + LPDDR5 64GB) は NVIDIA Ecosystem (CUDA / TensorRT / Jetpack SDK) 成熟度で competitor・Tenstorrent Blackhole ($1,500-$3,000・745 TOPS + GDDR6 32GB) はWorkstation-grade Higher tier。$400-$800 Mid-range tier で Axelera Metis が独自Architecture (D-IMC) + Energy efficiency差別化軸を持つ。
関連用語との違い
- vs Hailo-10H: Metisは214 TOPS + LPDDR4x 16GB + $400-$800・Hailo-10Hは40 TOPS + DDR4 8GB + $200-$400 (Mainstream Edge AI)
- vs NVIDIA Jetson AGX Orin: Metisは$400-$800 chip単体・JetsonはModule $1,999 + NVIDIA Ecosystem (CUDA/TensorRT)
- vs Tenstorrent Blackhole: Metisは214 TOPS Mid-range・Blackholeは745 TOPS Workstation-grade Higher tier
よくある質問（FAQ）
Q1: Digital In-Memory Computing (D-IMC) とは？ A: SRAM内のArray of bit-cellsで Computing performing (Multiply-Accumulate operations) 実行・Memory access overhead (Von Neumann bottleneck) を回避・10-30倍Energy efficiency improvement (vs Conventional GPU/CPU)。Axelera独自Patent技術・IBM Research出身Founder team の Academic technology background。
Q2: Hailo-10H vs Axelera Metis どちらを選ぶ？ A: Mainstream Edge AI Use case (Smart Camera / Drone / Robot・Power efficiency重視) → Hailo-10H ($350)・Edge AI Server / AI Workstation / Industrial Vision AI Mass scale → Axelera Metis ($600・5倍以上性能)。
Q3: Axelera AIブランド信頼性は？ A: 2021年設立・Founder = Fabrizio Del Maffeo (元IBM Watson AI Lead) + Evangelos Eleftheriou (元IBM Research・In-Memory Computing分野Industry Leader)・累計funding $120M+ + EU Government grants $40M ($160M total)・Netherlands Eindhoven government支援・Academic publications多数 + Patent portfolio強・新興だが Pioneering technology + Strong team。
まとめ
Axelera Metis AI = 2024年Q2 Axelera AI発表のEdge AI Server chip。214 TOPS INT8 + Digital In-Memory Computing (D-IMC) Architecture + LPDDR4x 16GB + 7-10W TDP + M.2/PCIe HHHL + $400-$800・Hailo-10H 5倍以上性能のEdge AI Server / AI Workstation向けMid-range tier。