ABS AI Workstation - 96GB Aggregate VRAM (3x 32GB AMD Radeon AI PRO R9700), AMD Ryzen Threadripper Pro 9965WX, 128GB ECC DDR5, 2x 2TB NVMe SSD, 2000W Gold PSU, Ubuntu (Zaurion Ruby ZRT9965WX-3XR9700)
- Made in USA
- TAA Compliant
- Operating System: Ubuntu
- CPU: AMD Ryzen Threadripper 9965WX (24-Core, 48-Threads)
- GPU: 3x AMD Radeon AI Pro R9700 (Total 96GB GDDR6)
- Motherboard: Gigabyte MH53-G40
- Memory: 128GB DDR5 Server RAM - ECC, Registered
- OS SSD: 2TB M.2 NVMe (Installed)
- Data SSD: 2TB M.2 NVMe (Installed)
- LAN: 2 x 10GB/s LAN, 1 x 10/100/1000 Mbps Management LAN
- PSU: Single 2000W ATX 80 PLUS Gold power supply
- Barebone: Gigabyte W773-H5D-AA01
Any questions? Ask CelesteSM will help you find out quickly.
Run Large Language Models Locally
Three AMD Radeon™ AI Pro R9700 GPUs with 96 GB total VRAM deliver high-performance local AI inference, RAG, coding assistants, and enterprise AI workloads on an open Ubuntu + ROCm™ stack.
Run LLMs Locally
Deploy 70B-class language models without the cloud. Large combined GPU memory enables longer context windows, faster inference, and secure on-prem deployment.
Open Software Stack
Ubuntu + AMD ROCm™ give you an open, standards-based platform for PyTorch, vLLM, and the broader open-source AI ecosystem — no vendor lock-in.
Private AI Infrastructure
Keep sensitive business data on-premises while cutting recurring cloud costs — purpose-built for organizations that require secure, offline AI computing.
Powered By an All-AMD Platform
Designed for Modern AI Workloads
Local LLM Inference
- ›70B-class language models
- ›AI chatbots
- ›Long-context inference
- ›Multi-user AI
AI Agents & RAG
- ›Enterprise knowledge base
- ›Document search
- ›AI Agents
- ›RAG pipelines
Private AI Platforms
- ›Open WebUI
- ›Ollama
- ›Onyx
- ›Private GPT
AI Inference Deployment
- ›OpenAI-compatible APIs
- ›Multi-model serving
- ›Production inference
- ›AI services
96 GB of combined GPU memory
Three AMD Radeon™ AI Pro R9700 GPUs provide up to 96 GB total GPU memory for large-model inference — handling larger models and longer context windows than a typical single-GPU workstation.
High-Performance Hardware
Why Choose This AI Workstation?
GPU Memory
Handle larger AI models and longer context windows than traditional single-GPU systems.
Open AI Stack
Ubuntu + AMD ROCm™ run PyTorch, vLLM, and open-source models with no vendor lock-in.
Enterprise Reliability
Server-grade memory, a workstation-class platform, and high-speed networking for continuous operation.
Warranty & Returns
Warranty, Returns, And Additional Information
Warranty
- Limited Warranty period (parts): 1 year
- Limited Warranty period (labor): 1 year
- Read full details
Return Policies
- Contact Manufacturer for Assistance
- This item is covered by Newegg.com's Manufacturer Warranty Only Return Policy.
- Read full return policy for details.
Manufacturer Contact Info
- Manufacturer Product Page
- Manufacturer Website
- Support Phone: 1-800-685-3471
- Support Email: customercare@abs.com
- Support Website
- View other products from ABS
The ABS ZRT9960X-3XR9700 is a tower AI workstation built for local model inference, GPU-accelerated development, and high-throughput compute workflows, making it well suited for AI engineers, research teams, and enterprise users who need 3 AMD Radeon AI Pro R9700 GPUs, 96GB total GDDR6, a 24-core Threadripper 9960X, and Ubuntu in a deployable desktop form factor.
Recommended For
- Teams building or serving local LLMs on Linux who want a single tower with high total VRAM for multi-GPU inference, model sharding, and larger context windows. AMD positions the Radeon AI PRO R9700 for local AI inference and memory-heavy workloads, and ROCm now officially supports the card on Linux.
- Developers and lab admins standardizing on Ubuntu/ROCm who prefer open-source AI stacks such as PyTorch, vLLM, Ollama, or llama.cpp over CUDA-only workflows. Real-world discussion around the R9700 centers on Linux-first deployment and iterative tuning rather than plug-and-play Windows use.
- Studios or technical users running memory-hungry image/video generation workflows locally, especially when 32GB-per-GPU capacity matters more than compact size. Community testing around ComfyUI and WAN workflows shows the R9700 being used specifically for larger local generation jobs on Ubuntu/ROCm.
- Engineering or enterprise buyers who need a desk-side AI workstation with workstation-class platform features such as ECC memory, 10GbE, remote management, and room for several dual-slot GPUs, instead of assembling a custom server from parts.
Spec Summary
| Attribute | Value |
|---|---|
| Best For | GPU-accelerated AI workloads and workstation-class compute — multi-GPU model training and inference, server-class ECC memory applications, and Linux-based development on a tower workstation platform |
| Processor And Core Count | AMD Ryzen Threadripper 9960X, 24 cores at 4.2GHz in a single-socket (sTR5) configuration |
| Cache And Chipset | L1 1920KB, L2 24MB, L3 128MB with AMD WRX90 chipset |
| Memory Configuration | 128GB DDR5 ECC RDIMM installed across 8 DIMM slots, 8-channel capable and expandable up to 1TB |
| GPU And VRAM | 3x AMD Radeon AI Pro R9700 GPUs totaling 96GB GDDR6; platform supports additional Gen5 dual-slot GPUs |
| Storage And Expansion | Two installed 2TB NVMe M.2 drives (OS and data), M.2 support with up to 4 M.2 slots (PCIe Gen5 x4) and front hot-swap support for 4 (plus optional 4) 3.5"/2.5" SATA bays |
| Expansion Slots And PCIe | 6 x PCIe Gen5 x16 expansion slots; system supports up to four dual-slot Gen5 GPUs in this chassis/platform |
| Wired Networking And Management | 2 x 10 Gb/s LAN (Broadcom BCM57416) plus 1 x 10/100/1000 Mb/s management LAN — total 3 RJ45 network ports including a management interface |
| Power | Single 2000W ATX 80 PLUS Gold power supply suitable for multi-GPU, high-power configurations |
| Operating System And Remote Management | Preinstalled with Ubuntu and includes a dedicated management LAN for remote administration |
| Model Identifiers | ABS ZRT9960X-3XR9700 tower workstation (GIGABYTE W773-H5D-AA01 GPU workstation platform referenced in features) |















