AI model training and deep learning
Eight GPUs with 768 GB of combined VRAM handle training and fine-tuning of large models on premises - your data never leaves your infrastructure, and costs do not grow with every experiment like cloud rentals do.
// 4U · GPU compute
AURA 4U-G8E2 - dual-EPYC GPU server, built to order
A production-grade 4U platform engineered around two AMD EPYC 9004/9005 processors and up to eight PCIe 5.0 GPUs for AI training, inference and HPC. Aura Digital configures every unit to your workload and delivers a personalized quote within 24 hours.
256CPU cores
8xPCIe 5.0 GPUs
6 TBDDR5 memory
PCIe 5.032 GT/s fabric
One dense node for your heaviest compute
The AURA 4U-G8E2 is a 4U rackmount server that combines two AMD EPYC processors, up to eight dual-slot PCIe 5.0 accelerators and up to 6 TB of DDR5 in a single 19-inch chassis. It is engineered for sustained, full-load operation, with fully redundant power, hot-swap cooling and IPMI remote management. Aura Digital, a Sofia-based systems integrator, assembles and validates each unit to order - you specify the GPUs, CPUs, memory and storage, and we deliver a build ready for the rack.
Accelerators
Up to 8 PCIe 5.0 GPUs in one node
The rear I/O carries nine full-height, full-length PCIe 5.0 x16 slots, hosting up to eight dual-slot GPUs for maximum accelerator density per rack unit. A special-order 10-GPU configuration is available via a PCIe switch backplane. The platform supports the current generation of NVIDIA data-center and pro-visualization cards alongside AMD Instinct.
- Up to 8 dual-slot PCIe 5.0 x16 GPUs; 10-GPU as special order via PCIe switch backplane
- NVIDIA RTX PRO 6000 Blackwell Server Edition (96GB), H200 NVL (141GB), H100, A100
- NVIDIA L40S (48GB), RTX 6000 Ada (48GB), RTX 5090 (32GB), RTX 4090 (24GB), plus AMD Instinct
- GPU-to-GPU interconnect on supporting models - NVIDIA NVLink Bridge or AMD Infinity Fabric Link
Processors
Dual AMD EPYC 9004/9005, up to 256 cores
Two AMD EPYC 9004 or 9005 processors on Socket SP5 deliver up to 256 cores in a single node, keeping high-bandwidth accelerators fed under sustained load. The platform supports processors with TDPs up to 500W per socket, so you can select for core count or per-core performance.
- 2x AMD EPYC 9004/9005 (Socket SP5, LGA 6096)
- Up to 256 cores total across both sockets
- Up to 500W TDP per socket supported
- Ample PCIe 5.0 lanes feed GPUs, NVMe and networking
Memory & storage
Up to 6 TB DDR5 with fast NVMe tiers
Twenty-four DIMM slots - 12 memory channels per CPU - scale to 6 TB of DDR5 RDIMM at up to 6400 MT/s, giving in-memory headroom for large models and datasets. On-board NVMe and hot-swap bays provide a fast local storage tier, with switch-based expansion when the workload needs more spindles.
- 24 DIMM slots, 12 channels per CPU, DDR5 RDIMM 5600/6000/6400
- Up to 6 TB total; module sizes 256/128/64/32 GB
- 2x M.2 (PCIe Gen3 x4, up to 22110) + 2x U.2 NVMe hot-swap (PCIe Gen4 x4)
- Optional PCIe 4.0 switch + 8x 2.5" U.2 NVMe hot-swap, or 8x 3.5" HDD bays
Thermal design
Push-pull cooling that prevents throttling
Six hot-swap fans form a push-pull airflow path tuned for a fully populated GPU chassis. Counter-rotating dual-motor central fans and high-static-pressure units drive air through optimally arranged heat sinks, while a rear exhaust module clears hot air - keeping accelerators off their thermal limits during long jobs.
- 3x 7200RPM counter-rotating dual-motor central fans (120x76mm, high static pressure)
- 3x 7200RPM high-static-pressure fans + rear exhaust hot-swap module (120x38mm)
- Push-pull configuration with intelligent airflow and optimized heat-sink arrangement
- Optional dual 40x28mm rear fan module for 10-GPU configurations
Power delivery
6400W redundant, Platinum efficiency
A 3+1 redundant supply built from four 1600W 80 PLUS Platinum CRPS modules delivers 6400W with the headroom a fully loaded GPU node demands, and stays online through a single PSU failure. Power reaches the components over an advanced PCB-based delivery design rather than a harness of cables.
- 6400W redundant, 3+1 x 1600W, 80 PLUS Platinum CRPS
- Front or rear AC power input to suit your rack layout
- PCB-based power delivery with 4oz/8oz thick copper
- Reduces cable clutter, improves airflow and cuts failure points
Remote management
Full lights-out control via IPMI
An IPMI-compliant baseboard controller with a dedicated management port gives you complete remote operation of the server. A browser-based console with HTML5 iKVM and remote media lets you install, reboot and recover the machine without ever standing in front of the rack.
- IPMI-compliant remote management with dedicated RJ45 port
- Browser-based graphical console with HTML5 iKVM and remote media
- Monitoring and control of Broadcom MegaRAID adapters
- Real-time server health monitoring
Reliability & signal integrity
Engineered for continuous production duty
Redundant power, hot-swap fans and hot-swap drives keep the node serviceable without downtime, all housed in a rigid chassis built for a fully loaded configuration. The board is laid out for signal integrity, so high-bandwidth links stay clean under load.
- Hot-swap cooling fans and drives, fully redundant power supplies
- Strong chassis engineered for dense GPU builds
- PCI-SIG compliant GPU backplanes with bottom-side MCIO connectors
- Shortest-possible trace lengths for low latency and maximum signal integrity
Networking & expansion
Onboard networking with room for fast NICs
Two Intel i350 1GbE ports plus a dedicated IPMI port cover management and baseline connectivity out of the box. A dedicated PCIe 5.0 x8 slot is reserved for a high-speed NIC or add-in card, so you can add the fabric your cluster needs without giving up a GPU slot.
- 2x 1GbE RJ45 (Intel i350) + dedicated RJ45 IPMI management port
- 1x PCIe 5.0 x8 (FH/HL) slot reserved for NIC or add-in card
- PCIe 5.0 (32 GT/s) throughout for GPUs, storage and networking
- Add high-speed Ethernet or InfiniBand adapters as a build option
01
Maximum performance
Dual EPYC processors and up to eight PCIe 5.0 accelerators deliver top-tier throughput for the most demanding AI and HPC workloads.
02
Warranty up to 5 years
Each build is backed by warranty coverage of up to five years for long-term peace of mind in production.
03
High compute density
Up to 256 cores, 8 GPUs and 6 TB of DDR5 packed into a single 4U node maximize compute per rack unit.
04
Superior cooling
A six-fan push-pull design with high-static-pressure fans keeps accelerators off their thermal limits during sustained jobs.
05
Cost-effective
Consolidating heavy compute into one dense, efficient node lowers your footprint, power overhead and operating cost.
06
Fast and flexible
Built to order and broadly configurable, so the platform is tailored to your exact workload and ready to scale.
Specifications
AURA 4U-G8E2 datasheet
Chassis & dimensions
- Form factor
- 4U rackmount, 19-inch
- Dimensions (D x W x H)
- 940 x 440 x 176.5 mm
- AC power input
- Front or rear
Processor
- CPU
- 2x AMD EPYC 9004/9005
- Socket
- Socket SP5 (LGA 6096)
- Cores
- Up to 256 cores total
- TDP
- Up to 500W per socket
Memory
- DIMM slots
- 24 (12 channels per CPU)
- Type
- DDR5 RDIMM 5600/6000/6400
- Max capacity
- Up to 6 TB
- Module sizes
- 256 / 128 / 64 / 32 GB
GPU & expansion
- GPUs
- Up to 8 dual-slot PCIe 5.0 x16 (10-GPU special order)
- Supported
- NVIDIA RTX PRO 6000 Blackwell SE, H200 NVL, H100, L40S, RTX 6000 Ada, RTX 5090, RTX 4090, A100, AMD Instinct
- Interconnect
- NVLink Bridge / Infinity Fabric Link - on supporting models
- Rear PCIe
- 9x PCIe 5.0 x16 (FH/FL) + 1x PCIe 5.0 x8 (FH/HL)
Storage
- M.2
- 2x socket (PCIe Gen3 x4, up to 22110)
- U.2 NVMe
- 2x hot-swap (PCIe Gen4 x4)
- Expansion
- Optional PCIe 4.0 switch + 8x 2.5" U.2 NVMe hot-swap, or 8x 3.5" HDD bays
Networking & management
- LAN
- 2x 1GbE RJ45 (Intel i350)
- Management port
- Dedicated RJ45 IPMI
- Remote console
- Browser-based, HTML5 iKVM with remote media
- RAID
- Broadcom MegaRAID monitoring & control
Cooling
- Fans
- 6x hot-swap, push-pull configuration
- Central
- 3x 7200RPM counter-rotating dual-motor (120x76mm)
- Static pressure
- 3x 7200RPM high-static-pressure fans
- Rear exhaust
- Hot-swap module (120x38mm); optional dual 40x28mm for 10-GPU
Power
- Capacity
- 6400W redundant (3+1 x 1600W)
- Efficiency
- 80 PLUS Platinum CRPS
- Delivery
- PCB-based, 4oz/8oz thick copper
Software & warranty
- OS / hypervisor
- Red Hat, Windows Server, VMware, SUSE, Ubuntu, Citrix
- Warranty
- Up to 5 years
- Sales model
- Built to order; personalized quote within 24h
Delivered builds
The first AURA 4U-G8E2 is already in production
We built and delivered the first machine for the NEOXIS GPU venture - from the pallet to the rack. The server is colocated in a Sofia data center and has been in operation since July 2026.
8× NVIDIA RTX PRO 6000 Blackwell Server Edition - 768 GB VRAM total2× AMD EPYC 9754 - 256 cores total24× Micron 64 GB DDR5 - 1.5 TB RAM2× Samsung 990 PRO 4 TB NVMe10GbE network card (2 ports)
Built for compute-intensive workloads
We configure the AURA 4U-G8E2 around your specific workload - these are the most common applications:
LLM inference and real-time AI services
Serve language models in real time - 96 GB of VRAM per card fits a large model without accuracy compromises, and the PCIe 5.0 fabric keeps responses fast even under many concurrent users.
Scientific computing and HPC
Simulations, computational chemistry, genomics, satellite-data processing - 256 cores and up to 6 TB of memory finish in days what workstations grind through for weeks.
3D rendering and video production
A render farm in a single chassis: eight GPUs split the frames between them - for animation studios, architectural visualization, VFX and color grading on heavy timelines.
Financial models and large-scale simulations
Monte Carlo simulations, risk analysis and backtests over years of market data - everything computes in memory, with no disk waits and no shared-cluster queues.
Private cloud and data-center infrastructure
The foundation of a private GPU cloud: virtualization with VMware or Proxmox, containers with Kubernetes, and GPU capacity shared between internal teams and clients.
Questions
Frequently asked
How many GPUs does the AURA 4U-G8E2 support?
The platform hosts up to eight dual-slot PCIe 5.0 x16 GPUs in its standard 4U chassis. A 10-GPU configuration is available as a special order via a PCIe switch backplane.
Which GPUs are supported?
Supported accelerators include the NVIDIA RTX PRO 6000 Blackwell Server Edition (96GB), H200 NVL (141GB), H100, A100, L40S (48GB), RTX 6000 Ada (48GB), RTX 5090 (32GB) and RTX 4090 (24GB), as well as AMD Instinct. Models with a direct GPU-to-GPU interconnect (e.g. H100, A100, AMD Instinct) can be linked via NVLink Bridge or Infinity Fabric Link; the RTX PRO 6000 Blackwell communicates over PCIe 5.0.
What CPUs and how much memory can it take?
It runs two AMD EPYC 9004/9005 processors on Socket SP5 for up to 256 cores total, with 24 DIMM slots supporting up to 6 TB of DDR5 RDIMM at speeds up to 6400 MT/s.
Is the power supply redundant?
Yes. The AURA 4U-G8E2 uses a 3+1 redundant configuration of four 1600W 80 PLUS Platinum CRPS modules for 6400W total, so the server keeps running through a single power-supply failure.
Can I manage the server remotely?
Yes. IPMI-compliant management with a dedicated port provides a browser-based console, HTML5 iKVM with remote media, Broadcom MegaRAID control and real-time health monitoring - full lights-out operation without physical access.
How do I get a price and what warranty is included?
Aura Digital builds every AURA 4U-G8E2 to order, so there are no fixed piecemeal prices. Configure your build and we deliver a personalized quote within 24 hours. Warranty coverage of up to five years is available.
Configure your AURA 4U-G8E2
Choose your GPUs, EPYC processors, memory and storage in the online configurator, and Aura Digital's engineers will return a personalized quote within 24 hours.