Two Flagships, Pooled.
Two NVIDIA RTX PRO 6000 MaxQ Blackwell cards — 96 GB of ECC GDDR7 each, 192 GB combined — on Intel's flagship Xeon W9-3575X server-board platform. The models no single card here can hold, the multi-model fleets, the whole-team GPU shares: this is the box they were waiting for.
96 + 96, Honestly Counted
GDDR7 with ECC — a full flagship-class card with its own memory, Tensor Cores, and MIG partitions.
Identical twin — run it pooled with GPU 1, or dedicate it to its own model, team, or tenant.
What Twin Flagships Change
Capacity, Isolation, and Density — Pick All Three
- 192 GB combined — split the largest open-weight models across both cards, or give long-context serving room no single card here offers.
- Perfect isolation — one card per production model means one model's traffic spike never starves the other.
- Two cards of Universal MIG — partition both flagships into isolated instances; a whole team shares safely.
- MaxQ density — the power-tuned editions are what let two flagship-class cards share one 2000 W 80 PLUS Gold supply, with headroom.
- ECC GDDR7 on both cards — weights on always-on services get bit-flip protection, twice over.
A Server Backbone Built to Feed Two
- Xeon W flagship silicon on the server-class Gigabyte MW83-RP0 board — the PCIe muscle two full-length flagships demand.
- 128 GB DDR5 ECC Registered — multi-day training and preprocessing pipelines finish without silent corruption.
- 2× 10 Gb/s LAN — two cards' worth of serving traffic, moved at wire speed; management LAN keeps IT out-of-band.
- 2 TB M.2 NVMe + 3.84 TB SSD — toolchains stay instant; datasets and checkpoints get their own lane.
- W773-W80 tower airflow — engineered to keep twin GPUs at full song around the clock.
Three Jobs Only a Duo Does
Beyond Any Single Card Here
Tensor and pipeline parallelism spread one massive model across both cards — 192 GB combined holds what this family's single-GPU towers cannot, with context to spare.
A Fleet of Two, Fully Isolated
Chat model on GPU 1, embedding or vision model on GPU 2 — each with a full 96 GB to itself, each isolated from the other's load, both served over dual 10 Gb/s LAN.
Twin Cards, Many Tenants
Universal MIG on both cards partitions the duo into isolated instances for researchers, services, and experiments — a departmental GPU utility in one tower, managed out-of-band by IT.
Scene photos are AI-generated illustrations of typical deployments, not photographs of this exact product.
Options Worthy of the Hardware
Pascari aiDAPTIV SSD
VRAM-expansion tier that offloads to flash — push large-model fine-tuning beyond even this duo's combined memory.
OpenClaw or NemoClaw
Enterprise AI agent framework pre-installed — agentic automation running the day the tower arrives.
Local LLM Package
Qwen 3.5 27B or higher, configured for private on-prem inference out of the box.
As Configured
| ZAW9-3575X-RP6000MQ · Zaurion Duo Aqua | |
|---|---|
| GPUs | 2× NVIDIA RTX PRO 6000 MaxQ Blackwell — 96 GB GDDR7 with ECC per card, 192 GB combined |
| CPU | Intel Xeon W9-3575X — workstation-class Xeon W platform, Socket on Gigabyte MW83-RP0 |
| Memory | 128 GB DDR5 ECC Registered server memory |
| Storage | 2 TB M.2 NVMe (OS) + 3.84 TB SATA SSD (data) |
| Network | 2× 10 Gb/s LAN + 10/100/1000 Mbps management LAN |
| Platform | Gigabyte MW83-RP0 board · W773-W80 tower barebone |
| Power | Single 2000 W ATX 80 PLUS Gold |
| OS | Ubuntu (optional OpenClaw / NemoClaw pre-installed; optional Qwen 3.5 27B+ package) |
Quick Answers Before You Buy
How does 192 GB combined actually work?
Each card carries its own 96 GB of GDDR7 with ECC. Modern stacks split one large model across both (tensor/pipeline parallelism), pool capacity for longer context, or dedicate one card per model.
Why MaxQ editions in a dual-GPU build?
Density and thermals: two power-tuned cards run side by side on a single 2000 W 80 PLUS Gold supply, with headroom, in a tower that lives outside a server room. Same 96 GB per card as the full-power edition.
Can it serve a whole team?
That's its natural habitat: two cards of Universal MIG partitions, dual 10 Gb/s LAN for wire-speed serving, and a management port so IT runs it like datacenter gear.
What software options can it ship with?
The Pascari aiDAPTIV SSD tier for VRAM expansion and fine-tuning, pre-installed OpenClaw or NemoClaw agent frameworks, and a Qwen 3.5 27B or higher local LLM package.
Government / public-sector eligible?
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. Build to Order — custom built and stress-tested after purchase.
