d-Matrix Corsair (DIMC Generative AI Chip・2024年)（ディーマトリックスコルセア）

概要

d-Matrix Corsair は、2024年Q4にUSA Santa Clara拠点のd-Matrix Inc. (2019年Sid Sheth (CEO・元Marvell Technology Senior Vice President 18年・Storage Networking) + Sudeep Bhoja (CTO・元Inphi / Marvell PHY R&D) 共同創業・累計funding $160M+ (Series B $44M 2022年 + Series C $110M 2024年Q1・Microsoft M12 / Triatomic Capital / Industry Ventures投資)) が発表したLLM Inference特化Generative AI Chip Corsair。Axelera Metis (D-IMC・SRAM内Computing) と類似のDigital In-Memory Computing (DIMC) Architecture (d-Matrix独自Patent技術・Sliced In-Memory Computing + Chiplet packaging) + 9,600 TOPS INT8 (Axelera Metis 214 TOPS の45倍・NVIDIA H100 PCIe 1,513 TOPS の6.3倍) + 2GB SRAM on-chip Computing (Memory bandwidth Bottleneck根本回避) + LPDDR5 256GB Card-level external memory + 600W TDP + PCIe Gen 5 x16 host接続・$3,000-$5,000/card (NVIDIA H100 PCIe $25k-$30k の 1/6-1/10) ・LLM Inference特化Chip (Training NPU support但・Inference optimized) ・Microsoft Azure採用予定 (2024年Q4発表・Project Forge LLM Inference Infrastructure integration・Microsoft Maia chip and pair deploy方針) ・NVIDIA H100直接競合のInference-Optimized AI Chip新興 Top startup。

主な特徴・仕組み

• Architecture: Digital In-Memory Computing (DIMC) ・SRAM内Computing・Memory bandwidth bottleneck回避
• Sliced In-Memory Computing: d-Matrix独自Patent技術・SRAM cell内computing + Chiplet packaging
• Performance: 9,600 TOPS INT8 + 4,800 TFLOPS BF16 / FP8 (NVIDIA H100 PCIe 1,513 TOPS INT8 + 756 TFLOPS BF16 vs 6.3倍 INT8)
• On-chip SRAM: 2GB SRAM (Memory bandwidth限界回避・5x more than typical GPU)
• External Memory: LPDDR5 256GB Card-level (Large LLM model fit・Mistral 8x22B / Llama 3 70B Q4 fit)
• Power: 600W TDP (NVIDIA H100 PCIe 350W vs 1.7倍だが性能6倍で総合効率優位)
• Form Factor: PCIe Gen 5 x16 Card (Full-Length Full-Height・Server insert)
• Supported Workload: LLM Inference特化 (Training NPU support but Inference optimized)
• Inference Performance: Llama 3 70B Q4 ~100 tok/s (NVIDIA H100 ~50 tok/s for 70B model vs 2倍高速)
• Aviator SDK: Compilation toolchain (PyTorch / ONNX / TensorRT-LLM compat)
• Microsoft Azure Adoption: 2024年Q4発表 Project Forge LLM Inference Infrastructure integration
• Use Cases: Hyperscaler LLM Inference / Multi-tenant AI Cloud / Cost-down LLM Inference Service
• Price: $3,000-$5,000/card (NVIDIA H100 PCIe $25k-$30k vs 1/6-1/10)

スペック比較表

具体例・対応製品

• d-Matrix Corsair PCIe Card: $3,000-$5,000・PCIe Gen 5 x16 + LPDDR5 256GB
• d-Matrix Aviator SDK: 無料・PyTorch / ONNX / TensorRT-LLM compat
• Microsoft Azure Project Forge integration: 2025年Q1-Q2 deploy予定・Azure LLM Inference Infrastructure
• d-Matrix Reference Designs: Hyperscaler LLM Inference / Multi-tenant AI Cloud
• 競合 NVIDIA H100 PCIe: $25k-$30k・1,513 TOPS + HBM3 80GB + 350W + NVIDIA Ecosystem深い (Customer Migration Cost大)

自作PCでの選び方・注意点

d-Matrix Corsair は「Hyperscaler LLM Inference Cost-down (NVIDIA H100 1/6-1/10価格)」「Microsoft Azure Project Forge LLM Inference Infrastructure integration」「Inference特化Workload (Training不要・Inference only deployment)」用途の本命LLM Inference-Optimized AI Chip。Consumer / 自作PC範囲外・カード単価 $3,000-$5,000で個人購入対象外・Hyperscaler / Cloud Provider / Large Enterprise customer専用。Digital In-Memory Computing (DIMC) Architecture (Axelera Metis類似Concept + 大規模Scale化) + 9,600 TOPS INT8 (NVIDIA H100 PCIe 6.3倍 INT8) + LPDDR5 256GB (H100 HBM3 80GB の 3.2倍 = Larger Model fit) + $3,000-$5,000価格 (H100 1/6-1/10) でLLM Inference Cost-down革新的提案。Microsoft Azure Project Forge (Microsoft独自LLM Inference Infrastructure・2024年Q4発表・Maia chipとpair deploy方針) 採用予定 (2024年Q4-2025年Q1 deploy timeline) で Hyperscaler proof実績期待。LLM Inference特化のため Training workload support限定的 (NVIDIA H100 / SambaNova SN40L等Training+Inference両用 vs differentiate)・Microsoft Azure / 他Hyperscaler (AWS / GCP / Oracle Cloud) のLLM Inference Service (GPT-4 / Claude / Llama等) コスト面 optimal化のChoice。Reddit / Hacker News / SemiAnalysis等の Tech media高 attention獲得 + Customer Adoption 2024年Q4-2025年が判断時期。

関連用語との違い

• vs NVIDIA H100 PCIe: Corsairは$3k-$5k + 9,600 TOPS Inference特化・H100は$25k-$30k + 1,513 TOPS + NVIDIA Ecosystem成熟
• vs Axelera Metis: CorsairはServer-grade 600W + LPDDR5 256GB + LLM特化・MetisはEdge AI 7-10W + LPDDR4x 16GB
• vs SambaNova Cardinal SN40L: CorsairはInference特化 Single-Card・SN40LはRack-level 256 chip Trillion-parameter

よくある質問（FAQ）

Q1: Microsoft Azure Project Forge とは？ A: 2024年Q4 Microsoft発表のAzure内部独自LLM Inference Infrastructure platform・Microsoft Maia 100 (Microsoft独自custom AI chip・2024年GA) + d-Matrix Corsair (External vendor chip) のpair deploy + Project Forge unified abstraction layer。Azure内LLM Inference workload (OpenAI GPT-4o / Microsoft Phi / Customer LLM) のCost-Performance optimization狙い。

Q2: NVIDIA H100 1/6-1/10価格でなぜ可能？ A: (1) DIMC Architecture = HBM3 expensive memory回避 + LPDDR5 cheaper memoryで cost-down + SRAM内Computing で Memory bandwidth limitation 根本解決・(2) Inference特化 design (Training NPU不要のため transistor budget効率) ・(3) New startup pricing strategy (Customer Adoption加速狙い) - 但しsupply constraint多 + Microsoft Azure exclusive deployment phase 2024年Q4-2025年。

Q3: d-Matrixブランド信頼性は？ A: 2019年設立・Founder Sid Sheth = Marvell Senior VP 18年 (Storage Networking分野industry veteran) + Sudeep Bhoja = Inphi/Marvell PHY R&D・累計funding $160M (Series C $110M 2024年Q1)・Microsoft M12 (Microsoft Venture Arm) Strategic Investor + Industry Ventures Top-tier・新興だが Founder credibility + Microsoft pre-commitment + d-Matrix Aviator SDKのCompilation toolchain完成度 で初期customer adoption phase進行中。

まとめ

d-Matrix Corsair = 2024年Q4 d-Matrix発表のDIMC Generative AI Chip。Digital In-Memory Computing Architecture + 9,600 TOPS INT8 + SRAM内Computing 2GB + LPDDR5 256GB + 600W + PCIe Gen 5 x16 + $3,000-$5,000・Microsoft Azure Project Forge採用予定 + Hyperscaler LLM Inference Cost-down (NVIDIA H100 1/6-1/10価格) + Inference特化のNVIDIA H100競合 Top startup。