A 12-core, 24-thread Intel Xeon keeps the accessible entry point into this GPU memory tier
affordable, without shrinking the GPU that matters. 72 GB of GDDR7 is
enough to develop, fine-tune, and run local AI models in the tens of billions of parameters
— right on the desk beside you. Kick off a fine-tuning run before a meeting, and it's
done by the time you're back — the kind of local iteration speed that turns
"let's check tomorrow" into "let's check right now."
72 GB Professional GPU128 GB DDR5Ubuntu ReadyAssembled in USA · TAA
72 GB
GDDR7 GPU memory
12c / 24t
Intel Xeon W5-2455X
128 GB
DDR5, ECC Registered
6 TB
NVMe SSD storage installed
2000 W
80 PLUS Gold PSU
AI Silicon Spotlight · Hardware First
What This Pairing Actually Buys You
An accessible CPU tier next to a memory-rich professional GPU, sized for real local AI work rather than benchmarks:
GPU · NVIDIA RTX PRO 5000 Blackwell
72 GB of GDDR7 for Local Model Work
72 GB of VRAM is enough headroom to develop against and run local models in the tens of billions of parameters without renting cloud GPU time.
Blackwell 5th-gen Tensor Cores with FP41 accelerate agentic and generative AI at architecture-native speed.
4th-gen RT Cores1 drive photoreal rendering previews without waiting on a render farm.
Universal MIG1 can partition the card into isolated instances when the workload calls for it.
CPU · Intel Xeon W5-2455X
12 Cores, 24 Threads, Exactly Enough
12 cores / 24 threads on LGA4677 keep data loading, tokenization, and orchestration moving ahead of the GPU, not behind it.
Workstation-grade ECC memory support brings server-class reliability to long training and inference runs.
The accessible tier of this platform — the same board also accepts far higher core-count CPUs when a workload actually needs them.
Frees the GPU to do GPU work — for GPU-bound inference and rendering, VRAM capacity matters more than CPU core count.
The Real Question
What 72 GB Actually Holds
GPU memory is the real ceiling on local model size. Here's the honest, illustrative math for what fits — not a benchmark, just the arithmetic:
FP16 · Full Precision
~30B
parameters
~2 bytes/parameter, plus headroom for context
INT8 · Quantized
~65B
parameters
~1 byte/parameter
INT4 · Quantized
~130B
parameters
~0.5 bytes/parameter, with some quality trade-off
Illustrative, approximate rule-of-thumb math (bytes-per-parameter × model size), not a vendor benchmark. Actual usable capacity depends on context length, batch size, and framework overhead.
AI Features · What This Actually Does
Six Jobs, One Accessible Tower
The listing's own AI Features, translated into what this specific configuration can actually do:
AI Development
With 72GB of GPU memory, iterate on 30B-class models at full precision, or up to roughly 70B quantized, entirely on this desk. Kick off a run, step away, and it's done — fast local development loops without a cloud API bill.
Data Science
12 cores / 24 threads keep data loading and preprocessing moving while the GPU handles the heavy lifting; 128GB of system memory holds large in-memory datasets.
HPC
PCIe Gen 5 bandwidth between CPU and GPU, plus Blackwell's mixed-precision Tensor Cores, keep simulation workloads fed with data instead of waiting on it.
AI-Driven Rendering & Graphics
72GB of VRAM holds large scene and texture data without the memory-driven quality trade-offs that stall lower-capacity cards, with RT Cores accelerating the preview.
Video Content & Streaming
Blackwell-generation video engines accelerate encode and decode for multi-stream editing, while 24 CPU threads handle timeline scrubbing and effects in parallel.
Game Development
The same 72GB GPU that runs local AI models also drives real-time ray-traced previews inside the engine editor — AI tooling and graphics work on one machine.
Scene photos are AI-generated illustrations of typical deployments, not photographs of this exact product.
Own vs. Rent
Local, Not Rented
A 72GB professional GPU on your own desk changes the economics and the guarantees of running AI workloads, not just the raw performance:
This Workstation
Cost per run$0 marginal
Your data & weightsStay on this desk
GPU accessAlways yours, on demand
Iteration limitHowever many you want
Typical Cloud GPU Rental
Cost per runBilled per token or per hour
Your data & weightsLeave your network
GPU accessShared — can queue
Iteration limitMetered by budget
Illustrative framing of the ownership trade-off, not a cost model for any specific cloud provider or workload.
OpenClaw agent framework (pre-installed); Qwen 3.5 27B+ local LLM package
FAQ
Quick Answers Before You Buy
What size of AI model can actually run on 72 GB of VRAM?
Illustrative rule-of-thumb math, not a vendor benchmark: roughly 30B parameters at full FP16 precision, roughly 65B at INT8, or roughly 130B at INT4. Actual usable capacity depends on context length, batch size, and framework overhead.
Why Xeon W5-2455X instead of a higher-core-count CPU?
It's the accessible entry point into this chassis lineup's 72GB GPU tier — 12 cores and 24 threads are enough to keep data loading, preprocessing, and orchestration ahead of the GPU, not behind it.
How much storage headroom is there?
Both of the platform's 2 M.2 slots are populated (2TB OS + 4TB data). The case's 4 installed SATA hot-swap bays (plus 4 more optional) are open, since this configuration is all-NVMe.
Does it come with an AI agent framework or local LLM pre-installed?
The AI agent framework (OpenClaw) is pre-installed; a Qwen 3.5 27B or higher local LLM package is an optional add-on configured at order time.
Is it eligible for government / public-sector purchasing?
It is TAA compliant and assembled in the USA, with compliance documents available for purchase review.
How is it supported?
1-year parts and labor limited warranty backed by ABS manufacturer support. As a Build to Order system, it is custom built and stress-tested after purchase.
Assembled in USA
TAA Compliant
72 GB Professional GPU
Build to Order flexibility
1. Intel-published core/thread count for the Xeon W5-2455X. GPU architecture features (5th-gen Tensor Cores, 4th-gen RT Cores, Universal MIG) are NVIDIA-published for the Blackwell professional GPU architecture generation, not SKU-specific benchmark figures.
ABS Zaurion Ruby ZAW5-2455X-RP500072 · Gigabyte W773-W80 platform (LGA4677 socket) · Optional components and software are configured at time of purchase.
Learn more about the ABS ZAW5-2455X-RP500072
Model
Brand
ABS
Series
Zaurion Ruby
Model
ZAW5-2455X-RP500072
Part Number
ZAW5-2455X-RP500072
Details
Workload
AI Workstation
Type
Tower
AI Features
AI Development: Accelerate AI development and inference workloads, and build agentic AI applications. Data Science: Accelerate your data pipelines. HPC: Accelerate scientific breakthroughs with enhanced accuracy. AI-Driven Rendering and Graphics: Deliver stunningly realistic designs and simulations with speed and precision. Video Content and Streaming: Boost performance for live media and video pipelines. Game Development: Drive new possibilities for game development with AI.
Processor
CPU Type
Intel Xeon W5-2455X
CPU Socket
LGA 4677
Number of CPU Cores
12
# of CPU Threads
24
# of Processors Installed
1
MAX Processors
1
Chipset
Intel W790
Memory
Installed Memory Size
128GB
Memory Type
DDR5
Memory Features
ECC
Number of Memory Slots
8
Storage
SSD (Installed)
OS SSD: 2TB M.2 Data SSD: 4TB M.2
Graphic Controller
Servers_Graphic Installed
1x NVIDIA RTX Pro 5000 Blackwell 72GB
GPU Type
NVIDIA RTX Pro 5000 Blackwell
Video Memory
72GB
Networking
LAN Ports
2 x 10 Gb/s LAN 1 x 10/100/1000 Mb/s Management LAN
Software
Operating System
Ubuntu
Power
Power Characteristics
Single 2000W ATX 80 PLUS Gold power supply
Energy-Efficient
2000W
Features
Features
GIGABYTE W773-W80 GPU Workstation
Additional Information
First Listed on Newegg
August 20, 2026
Customer Reviews of the ABS ZAW5-2455X-RP500072
There are no reviews yet.
Do you own this product?
Warranty & Returns
Warranty, Returns, And Additional Information
Warranty
Limited Warranty period (parts): 1 year
Limited Warranty period (labor): 1 year
Return Policies
Return for refund within: non-refundable
Return for replacement within: non-replaceable
This item is covered by NeweggBusiness.com's Manufacturer Only Return Policy.
