Marvell Octeon 10 Fusion (DPU + ARM Server CPU・2024年)（マーベルオクテオンテンフュージョン）

概要

Marvell Octeon 10 Fusion は、2024年Marvell Technology Inc. (米Santa Clara本社・1995年Sehat Sutardja + Pantas Sutardja兄弟設立・半導体IC設計大手・累計売上$6B/year・Custom silicon + Networking specialty + Storage controllers + Wireless + Custom ASIC + DPU/Smart NIC market leadership・米半導体industry 中規模major player) が発表したOcteon 10 Fusion・Marvell Octeon family DPU (Data Processing Unit) + ARM Server CPU Fusion product・Marvell Custom silicon + Networking specialty heritage継承 + ARM Server CPU + DPU + AI accelerator + Networking ASIC integrated platform・Networking + Compute Fusion approach (Datacenter compute + Networking + Storage offload integrated single chip)。前世代Octeon 10 2022年 32-core ARM Neoverse N2 DPU + AI accelerator integration (前世代Octeon 10 base + AI accelerator + Smart NIC + Limited compute + DPU positioning) → Octeon 10 Fusion 2024年 evolution = 96 cores Neoverse N2 (3倍 cores expansion・DPU-centric → Compute-capable evolution + Smart NIC + ARM Server CPU dual-purpose) + DPU integration (Smart NIC offload + Packet processing + Security crypto IPsec / TLS / Disk encryption + AI inference accelerator) + 1 Tbps Networking throughput (Industry-leading DPU networking throughput・NVIDIA BlueField-3 400Gbps NDR vs Marvell Octeon 10 Fusion 1 Tbps = Industry-leading・Telecom + 5G core + Multi-port networking + High aggregate bandwidth) + TSMC 5nm process + Telecom (Cellular base station offload + 5G core function offload + Network slicing + Telecom edge compute) + Datacenter Smart NIC (Storage + Network virtualization offload + Hypervisor offload) + Networking ASIC alternative (前世代 separate Networking ASIC + Server CPU + Smart NIC discrete design vs Octeon 10 Fusion = Integrated single chip + Cost reduction + Power reduction)・Networking + Compute Fusion approach + Telecom + Datacenter edge compute use case alignment。

主な特徴・仕組み

• Vendor: Marvell Technology Inc. (1995年設立・半導体IC設計大手・累計売上$6B/year)
• Cores: 96 cores ARM Neoverse N2 (前世代Octeon 10 32-core vs 3倍 cores expansion)
• Architecture: ARM v9.0a (SVE2 + MTE + CCA Confidential computing)
• Process: TSMC 5nm
• Networking: 1 Tbps throughput Industry-leading DPU networking
• DPU Features: Smart NIC offload + Packet processing + Security crypto (IPsec + TLS + Disk encryption)
• AI Accelerator: Integrated AI inference accelerator (Octeon 10 series heritage継続)
• Use Cases: Telecom (5G core + Cellular base station + Network slicing) + Datacenter Smart NIC + Networking ASIC alternative
• Predecessor: Octeon 10 2022年 32-core (DPU-centric) → Octeon 10 Fusion 2024年 96-core (Compute-capable)
• Strategic Position: Networking + Compute Fusion approach + Datacenter Smart NIC + Telecom edge compute

スペック比較表

具体例・対応製品

• Marvell Octeon 10 Fusion (Standard): $3,000-$5,000・96 cores + 1 Tbps + DPU integrated
• Marvell Octeon 10 (前世代・2022年): 32 cores Neoverse N2 + DPU + AI accelerator
• Marvell Prestera (Networking ASIC・別family): Data Center Networking ASIC family
• Marvell Storage Controllers: Storage + Custom ASIC family
• 競合 NVIDIA BlueField-3: 400Gbps NDR + 16 ARM cores + AI Datacenter Smart NIC dominant
• 競合 AMD Pensando DPU: 200Gbps + 32 ARM + Datacenter Smart NIC

自作PCでの選び方・注意点

Marvell Octeon 10 Fusion は「Telecom (5G core + Cellular base station offload + Network slicing + Telecom edge compute・5G infrastructure requirements + Telecom OEM customers Ericsson + Nokia + Samsung Networks + ZTE等 deployment)」「Datacenter Smart NIC (Storage + Network virtualization offload + Hypervisor offload + Datacenter operating cost reduction)」「Networking ASIC alternative (前世代 separate Networking ASIC + Server CPU + Smart NIC discrete design vs Octeon 10 Fusion = Integrated single chip + Cost reduction + Power reduction + Board space reduction)」用途のNetworking + Compute Fusion approach DPU + ARM Server CPU Fusion product。Marvell Octeon family DPU heritage継承 + ARM Server CPU + DPU + AI accelerator + Networking ASIC integrated platform で Networking + Compute Fusion approach (Datacenter compute + Networking + Storage offload integrated single chip)・前世代Octeon 10 2022年 32-core ARM Neoverse N2 DPU + AI accelerator integration → Octeon 10 Fusion 2024年 evolution 96 cores expansion 3倍 + DPU-centric → Compute-capable evolution + Smart NIC + ARM Server CPU dual-purpose。1 Tbps Networking throughput (Industry-leading DPU networking throughput・NVIDIA BlueField-3 400Gbps NDR vs Marvell Octeon 10 Fusion 1 Tbps = Industry-leading・Telecom + 5G core + Multi-port networking + High aggregate bandwidth) で Telecom (Cellular base station offload + 5G core function offload + Network slicing + Telecom edge compute) + Datacenter Smart NIC (Storage + Network virtualization offload + Hypervisor offload) high networking throughput requirements alignment。Marvell Custom silicon + Networking specialty heritage継承 (1995年設立・29年 半導体IC設計大手・累計売上$6B/year・Custom silicon + Networking specialty + Storage controllers + Wireless + Custom ASIC + DPU/Smart NIC market leadership・米半導体industry 中規模major player) で Industry-leading DPU + Networking specialty trust + Telecom OEM customers + Datacenter customers diverse customer base + Long-term silicon design capability。Networking + Compute Fusion approach (Networking ASIC + Server CPU + Smart NIC discrete design integrated single chip + Cost reduction + Power reduction + Board space reduction) で Telecom + Datacenter edge compute use case optimization + Cost-effective deployment + Compact form factor。但しIndustry-leading DPU networking (NVIDIA BlueField-3 dominant AI Datacenter Smart NIC) vs Marvell Octeon 10 Fusion = Different market positioning (Telecom + Datacenter Smart NIC alternative + Networking + Compute Fusion approach) + Marvell-specific use case strength (Telecom edge compute + Networking ASIC alternative)。

関連用語との違い

• vs NVIDIA BlueField-3 DPU: MarvellはNetworking 1 Tbps + Compute 96 cores Fusion approach・BlueFieldはAI Datacenter Smart NIC 400Gbps + 16 ARM cores
• vs AMD Pensando DPU: MarvellはNetworking + Compute Fusion 96 cores・PensandoはDatacenter Smart NIC + 32 ARM
• vs Microsoft Cobalt 100 / Google Axion: MarvellはDPU + ARM CPU Fusion + Telecom focus・Cobalt/AxionはHyperscaler internal ARM Server CPU + General compute

よくある質問（FAQ）

Q1: Networking + Compute Fusion approach 意義は？ A: 前世代 separate Networking ASIC (Marvell Prestera等) + Server CPU + Smart NIC discrete design (Multiple chips + Higher cost + Higher power + Larger board space) vs Octeon 10 Fusion = Integrated single chip + Cost reduction + Power reduction + Board space reduction (Total cost of ownership reduction + Telecom + Datacenter operating cost optimization + Edge compute compact form factor)・Networking + Compute Fusion approach + Marvell-specific differentiated positioning。

Q2: 1 Tbps Industry-leading networking vs BlueField-3 400Gbps 比較は？ A: NVIDIA BlueField-3 DPU 400Gbps NDR (AI Datacenter Smart NIC dominant + InfiniBand integration + GPU integration) vs Marvell Octeon 10 Fusion 1 Tbps (Industry-leading aggregate bandwidth + Multi-port networking + Telecom + 5G core high throughput requirements) → 各target market alignment・BlueField AI Datacenter focus + Marvell Telecom + Datacenter Smart NIC + Networking ASIC alternative + Higher aggregate bandwidth required market focus + Different value proposition。

Q3: Telecom + Datacenter dual market positioning は？ A: Telecom market (5G core + Cellular base station offload + Network slicing + Telecom edge compute・Telecom OEM customers Ericsson + Nokia + Samsung Networks + ZTE + 多Telecom infrastructure customers) + Datacenter Smart NIC market (Storage + Network virtualization offload + Hypervisor offload + Datacenter operating cost reduction) dual market positioning・Multi-market value proposition + Diverse customer base + Marvell Networking specialty heritage strength + Long-term silicon design ROI + Industry-wide DPU + ARM CPU Fusion approach evolution。

まとめ

Marvell Octeon 10 Fusion = 2024年Marvell発表のDPU + ARM Server CPU Fusion product。96 cores Neoverse N2 + DPU integration + 1 Tbps Industry-leading networking + TSMC 5nm + Smart NIC offload + Packet processing + Security crypto + AI accelerator + Telecom + Datacenter Smart NIC + Networking ASIC alternative + Networking + Compute Fusion approach + Marvell Networking specialty heritage継承。