KOㅤ|ㅤEN

dg5R

Datacenter in a Box

Overview

Independent AI Infrastructure Managing 10kW Ultra-High Heat

dg5R is equipped with the most advanced deepgadget liquid cooling technology, enabling up to 10 AI chips to operate at room temperature. With comprehensive cooling covering CPU, memory, and NICs, performance is pushed to its limits, while the latest architecture support and remote monitoring maximize operational convenience. dg5R is the most practical AI infrastructure available now, from challenging server rooms to small and large data centers.

Recommended Workloads

What You Can Do Right Now with dg5R

Inference for Frontier-Scale Open LLMs

With 33,410 TFLOPS FP8 and 1.41TB of HBM3e memory, dg5R can accommodate frontier-scale open model inference on a single server, including Llama 3.1 405B, Qwen3-235B-A22B, and DeepSeek-class models. For 70B-class models, the system is well suited to higher concurrency and more stable production serving.
RAG-powered knowledge search, enterprise chatbots, and Copilot services
Document summarization, classification, Q&A, and report draft automation
High-throughput batch inference and concurrent user traffic handling

AI Inference

Domain-Specific Model Tuning and Validation

Backed by 8,350 TFLOPS BF16 performance and 1.41TB of HBM3e memory, dg5R enables LoRA- and SFT-based tuning and evaluation of 70B-class LLMs within a single server. Training, validation, inference, and deployment can be iterated in one place to shorten the experimentation cycle.
Domain-specific model adaptation with LoRA and SFT
Iterative experimentation with checkpoints, evaluation, and inference outputs
Private AI assistants tailored to internal policy, terminology, and tone

LLM Tuning

Industrial & Financial High-Performance Computing

Accelerated Simulation for Industry and Finance

With 300 TFLOPS FP64 performance and 1.41TB of HBM3e memory, dg5R accelerates large-scale numerical workloads on a single server, including FEA, CFD, and Monte Carlo risk simulation.
Large-scale FEA structural analysis and CFD simulation
Parallel computation for Monte Carlo risk and quantitative finance models
Digital twin and infrastructure simulation workflows
Hybrid pipelines combining AI inference with numerical computing

Industrial & Financial High-Performance Computing

Workflows for Image Generation, Transformation, and Visualization

With 3,545 TFLOPS of RT Core performance, 9,360 TFLOPS FP16 throughput, and 960GB of GDDR7 memory, dg5R can accelerate large-scale image generation, image transformation, 3D rendering, and visualization workflows on a single server.
Batch inference and high-volume generation with large image models
Image transformation, upscaling, style transfer, and post-processing pipelines
3D rendering, digital content creation, and visualization workloads
Extended visual AI workflows built on private brand or domain data

Image Workflows

Highlights

AI Power in a Single Server

Max AI Processors

10GPUs

LLM Inference

33.4PFLOPS

FP8 기준

LLM Training

8.3PFLOPS

BF16 기준

High-Speed Network

4Tb/s

Up to 10x NIC configuration

dg5R features a total of 15 PCIe 5.0 x16 slots, with up to 10 available for AI processor expansion. LLM inference theoretical peak is 33.4 PFLOPS with 10 NVIDIA H200 NVL GPUs under FP8 Tensor and Structured Sparsity conditions; LLM training theoretical peak is 8.3 PFLOPS under BF16 Dense with the same configuration. For networking, multiple 400Gbps+ InfiniBand / RoCE NICs can be installed, supporting both InfiniBand and Ethernet protocols.

에너지 효율

1.41PUE

공랭식 1.85 대비 약 23.8% 개선

공조 비용 절감

52.4%

서버 내부 냉각 절감

54.0%

전체 전력 비용 절감

23.8%

공랭 vs dg5R

	공냉 서버	dg5R
IT 전력	672.0 kW	672.0 kW
서버 내부 냉각	300.0 kW	132.8 kW
중앙 공조	224.0 kW	96.0 kW
UPS 손실	33.6 kW	33.6 kW
기타	15.0 kW	10.0 kW
총 전력	1,244.6 kW	944.4 kW
PUE	1.85	1.41

* Based on 100-server datacenter (IT power 672 kW), same workload. Air-cooled servers require ~20°C intake for stable operation; figures calculated under that condition. dg5R (self-contained L2A) figures calculated at 35°C intake.

Supports Up to 10 Latest AI Accelerators

Supports up to 10 of the latest NVIDIA GPUs as well as a wide range of next-generation AI accelerators from AMD, Tenstorrent, FuriosaAI, and more. Configure the perfect environment for your workloads.

High-Performance Enterprise

NVIDIA H200 NVL

Large-Scale Inference

NVIDIA RTX PRO 6000

Best Value

NVIDIA RTX 5090

Technology

More Thorough and Reliable deepgadget Liquid Cooling

Scientific L2A (AALC) Cooling Flow with CFD Analysis

AALC(Air-Assisted Liquid Cooling) is a structure where liquid captures heat and air manages final heat rejection. Cold plates in direct contact with major heat sources such as CPUs, AI processors, and NICs absorb heat into the coolant, which then passes through built-in radiators to release that heat into the air. The discharged heat is handled by the data center’s existing HVAC system, allowing flexible deployment within conventional air-cooled infrastructure without direct facility water connections. The entire cooling path is optimized through CFD-based design to deliver high cooling efficiency and stable operation.

Liquid Cooling for AI Chips, CPU, NIC, and Memory

Liquid cooling is applied to all heat-generating components including NIC, chipset, and CPU, Memory not just GPUs. Efficiently manages heat across the entire system to maintain optimal performance.

Patented StackFlow Technology

deepgadget's patented StackFlow technology achieves maximum heat dissipation efficiency in limited space. Stacked radiator structure, flow path design, and parallel cooling flow enable stable temperature management.

• Cooling System for Server — Patent No. 10-2879706

In-house Cooling Plate Design Capability

We design cooling plates optimized for each chip including GPU, CPU, and NIC to maximize heat transfer efficiency. Patented design ensures uniform cooling performance.

• Water Cooling Device for Computer and Driving Method — Patent No. 10-2118786

Powerful Flow Rate from 4 Pumps

Four pumps — two per loop — continuously push coolant through two independent cooling circuits. With up to 16 liters per minute circulating at full force, heat from AI processors and CPUs is absorbed and carried away the instant it's generated. And if one loop encounters an issue, the other keeps cooling without interruption, so the server never has to throttle down.

Certified Quality & Trust

Verified by domestic and international certification bodies — ISO 9001 quality management, direct in-house production, and venture company recognition for technology and growth potential.

Quality Management

ISO 9001

Ministry of SMEs

Direct Production Cert.

Ministry of SMEs

Venture Company Cert.

Configuration

Latest Architecture for Large-Scale Workloads

Latest AI Processor Support

NVIDIAPRO 6000 Blackwell · H200 NVL · RTX 5090
AMDRX 9070 XT · AI PRO R9700
TenstorrentBlackhole p150a · Wormhole n300s
FuriosaAIRNGD · WARBOY

Latest CPU Support (Dual)

Intel Xeon 6700 Series
Intel Xeon 6900 Series
AMD EPYC 9005 Series
Dual CPU Configuration Support

PCIe Gen 5 Slots

Up to 15 PCIe 5.0 x16 Slots
Up to 128 GB/s bidirectional bandwidth per slot

DDR5-6400 Memory Support

ECC RDIMM Support
Intel Xeon 6700 SeriesUp to 32 DIMMs (4TB)
Intel Xeon 6900 Series/AMD EPYC 9005 SeriesUp to 24 DIMMs (3TB)

High-Speed NIC Support

NVIDIA ConnectX-7 NDR / 400GbE
NVIDIA ConnectX-6 HDR / 200GbE
InfiniBand / Ethernet Dual-Mode
RDMA, RoCEv2 Support

1GbE & IPMI Support

Comes standard with 2 x 1GbE Base-T ports and IPMI remote management. Monitor and manage your server reliably from anywhere.

Expandable Storage

Supports M.2 NVMe SSD (HW RAID) and up to 16 U.2 NVMe SSDs. Capacities range from 3.84TB to 61.44TB per drive; with 61.44TB drives, total storage reaches up to 983TB.

Enterprise 3200W Hot-Swap Redundant PSU

Equipped with four 3,200W 80 Plus Titanium certified PSUs. Supports both hot-swap and redundancy for stable power delivery during live server operation.

dg5R Supporteddg5R

Compare Against

NVIDIA RTX PRO 6000 Blackwell SE

NVIDIA H100 NVL

Leads96GB GDDR7

MemoryGB

94GB HBM3

234

TF32TFLOPS

494Leads

468

BF16TFLOPS

989Leads

936

FP16TFLOPS

1,979Leads

3,784

FP8TFLOPS

3,958Leads

1.88

FP64TFLOPS

34Leads

234

FP32TFLOPS

494Leads

Leads354.5

RT CoreTFLOPS

—

* Performance figures are based on manufacturer-disclosed specifications; some are estimated. Actual performance may vary depending on system configuration and environment.

Support

Convenient Monitoring, Proactive All-in-One Management.

Dashboard

gadgetini, our web-based monitoring software, covers the entire server across six dedicated panels: Overview, Cooling Health, Compute Health, Network Health, History, and Alerting. Color-coded Normal / Warning / Critical status gives you an instant read on server health, while real-time data spans per-loop coolant temperature, leakage, and level — alongside AI processor and CPU temperature, power, and utilization — plus NIC link status and InfiniBand chipset temps. A 24-hour history graph lets you trace anomalies back to their source, and threshold breaches trigger instant email and Slack alerts. Accessible from desktop or mobile, anywhere.

01 / 06

Overview

Server status is color-coded Normal / Warning / Critical at a glance. Cooling (Inlet/Outlet/ΔT, leakage, level), AI processors, CPUs, and network link status are all summarized on one panel.

Display

The built-in front LCD display lets you check server status directly. It shows all information available on the dashboard, and display modes can be freely selected via the gadgetini web UI.

Q: What happens if there is a leak in the Liquid Cooling System?

deep gadget is engineered as a fully sealed structure, minimizing the risk of leakage under normal circumstances. However, we do have precautionary measures in place. Our system is equipped to shut down automatically in the event of a critical failure such as a leak caused by physical shock or other factors. Additionally, the amount of coolant that may leak due to our liquid cooling system design is minimal, typically not exceeding tens of milliliters. This level of leakage has negligible impact on other devices and minimal effect on deep gadget itself. In the event of any malfunction, our engineers will promptly address the issue, replacing the problematic part and reinstalling it as needed.

Instant deep learning.
Preloaded for AI. Ready to train from day one.

deep gadget comes preloaded with all the software needed for AI research and development—so you can start deep learning the moment you power it on. From the OS to the deep learning stack, everything is optimized by the deep gadget team. Designed for effortless compatibility, so you can stay focused on your research.

OS

Ubuntu

RHEL

Rocky Linux

Windows

Drivers

NVIDIA

AMD

Tenstorrent

Infiniband/GbE

Runtime & Libraries

CUDA

cuDNN

NCCL

cuBLAS

TensorRT

ROCM

Deep Learning Frameworks & Solutions

PyTorch

TensorFlow

DeepSpeed

Ollama

Horovod

vLLM

3rd Party Solutions

Development Tools

MPI

Anaconda

Docker

Specifications

dg5R Specifications

Noise

30 ~ 62 dB (Full load)

Form Factor

447mm x 911.5mm x 266mm(6U)

CPU

Intel Xeon 6700 Series — Dual, up to 144 Cores/CPU
6725P · 6737P · 6747P · 6761P · 6756E · 6780E
Intel Xeon 6900 Series — Dual, up to 128 Cores/CPU
6952P · 6972P · 6960P · 6978P · 6980P
AMD EPYC 9005 Series — Dual, up to 128 Cores/CPU
9135 · 9355 · 9455 · 9555 · 9755

GPU

Up to 10x
NVIDIA RTX PRO 6000 Blackwell Server Edition
NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition
NVIDIA H200 NVL Tensor Core
NVIDIA RTX 5090
AMD RX 9070 XT · AMD AI PRO R9700
tenstorrent Wormhole n300s · tenstorrent Blackhole p150a
FuriosaAI RNGD · FuriosaAI WARBOY

Memory

DDR5-6400 RDIMM (ECC)
Intel 6700: up to 4,096 GB (32 × 128 GB)
Intel 6900 · AMD: up to 3,072 GB (24 × 128 GB)

Primary Storage

4× M.2 NVMe SSD (1 / 2 / 4 TB) HW RAID Controller included
Up to 16× U.2 NVMe SSD (up to 983 TB)

PCIe

Up to 10 Gen 5 x16 slots

PSU

4 × Hot swappablle redundant 3,200 W 80 PLUS Titanium

Networking | Max Bandwidth

2 x 1GbE Base-T, IPMI
NVIDIA ConnectX-6 — InfiniBand EDR / HDR200 / Ethernet 200GbE
NVIDIA ConnectX-7 — InfiniBand NDR / Ethernet 400GbE

Supported OS

Ubuntu · RHEL · Rocky Linux · Windows Server

Warranty

3-Year Warranty

dg5W

Rack / Tower type