TN 4208-K26 Dual-Socket AI Training & Inference Server

TN 4208-K26 is a 4U dual-socket AI training & inference server with NPU accelerators and Kunpeng 920 processors. Outstanding computing performance, balanced configuration, efficient deployment. Optimized for billion-parameter LLM inference/training. Supports private deployment and data center deployment.

• Accelerators: up to 8×full-height full-length single-width NPU cards • CPU: 2×Kunpeng 920 (48C/64C), 2.6GHz • Memory: 32×DDR4 RDIMM, max 4TB • Storage: 12×3.5" SAS/SATA hot-swap (4 NVMe-capable) + 2×M.2 SSD • RAID: 0/1/10/5/50/6/60 with super-capacitor cache protection • PCIe: up to 11 slots (8 dedicated for AI accelerators) • Network: 2×half-height NICs + 2×Flex IO slots • Fans: 6/8 hot-swap N+1 redundant • Management: BMC dedicated GE NIC, Redfish/SNMP/IPMI 2.0 • PSU: 4×2000/2600/3000W AC Platinum/Titanium 2+2 or 3+1 redundant • OS: BC-Linux / Ubuntu / CentOS / openEuler / UOS / Kylin

01 Leading Architecture, Outstanding Performance Up to 8×full-height NPU cards with 2×Kunpeng 920 (48C/64C). CPU+NPU collaborative computing optimized for billion-parameter LLM inference/training. 32×DDR4 RDIMM up to 4TB. 12×3.5" SAS/SATA (4 NVMe-capable) + 2×M.2 SSD. 02 Flexible Expansion, Reliable Up to 11 PCIe slots (8 dedicated for AI accelerators) + 2×Flex IO slots for 25/100GE NICs. Up to 4×3000W Titanium PSU with 2+2 or 3+1 redundancy. Super-capacitor RAID cache protection. 03 Intelligent Management iBMC auto version push and contactless upgrade, 20x faster than manual. 75% tooling for pipeline deployment, 10x deployment efficiency. Asset auto-identification with real-time tracking for 100% inventory accuracy. Redfish/SNMP/IPMI 2.0 open APIs. 04 Use Cases Kunpeng + NPU collaborative computing for smart city, government cloud, public security, education, internet, AI, telecom, finance. Workloads: internet, big data, distributed storage, HPC, databases, cloud, SDS, AI training/inference.

Front • USB 3.0 • VGA • 12×3.5" hot-swap drive bays • Status LEDs / ID button Rear • USB 3.0 • VGA • 1×BMC GE dedicated NIC • 2×Flex IO slots • 2×half-height NIC slots • 8×AI accelerator PCIe slots • Regular PCIe slots • 4×PSU connectors

• 8×NPU + dual Kunpeng 920 collaborative computing • Optimized for billion-parameter LLM inference/training • 4TB memory + 12 bays (4×NVMe) for training data throughput • 11×PCIe slots for I/O expansion • 4×3000W Titanium PSU with 2+2 or 3+1 redundancy • Super-capacitor cache protection against power loss • iBMC intelligent O&M: auto push, contactless upgrade • Open APIs: Redfish / SNMP / IPMI 2.0

AI LLM training (billion-parameter), AI inference, smart city video analytics, government cloud AI, public security face recognition, education AI platforms, internet recommendation/search, telecom AI NFV, financial risk modeling, medical AI imaging

• Form: 4U AI server 175×447×898mm (H×W×D) • Operating: 5-35°C (ASHRAE A1/A2) • Humidity: 8-90% non-condensing • PSU: 4×2000/2600/3000W AC Platinum/Titanium, 2+2 or 3+1 redundant • Fans: 6/8 hot-swap modules, N+1 redundant • Management: BMC dedicated GE NIC, Redfish/SNMP/IPMI 2.0

Standard: TN 4208-K26 + 2×920(64C) + 8×NPU + 512G + 3000W Titanium + Kylin Fully customizable Contact: Tunan (Shenzhen) Information Technology