Eight cards, one chassis
Eight NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs in a single 4U rackmount — 768 GB of GDDR7 across PCIe Gen5, enough headroom to serve 70B-class models to 100–500 concurrent users without leaving the rack.
EMARQUE built
EMARQUE AI Server
8× NVIDIA RTX PRO 6000 Blackwell Server Edition (96 GB GDDR7 each) in a 4U PCIe rackmount reference platform. Available through EMARQUE.
768GB GDDR7 max
8PCIe Gen5 cards
4UAir-cooled
Key features
Configuration overview.
Manufacturer-defined features from the published datasheet.
Datacentre-grade chassis, EMARQUE delivered
Giga Computing G493-class reference platform — the same hardware that ships into hyperscaler labs — supplied through EMARQUE with local commissioning, warranty handling, and Tier-1 support in Malaysia.
Built to run multi-tenant
MIG-style partitioning per card lets you carve the server into isolated inference workloads — separate teams, separate models, separate quotas on shared hardware. No noisy-neighbour fights, full observability via the BMC.
Storage and network that match the GPUs
Up to 60 TB U.2 NVMe on PCIe Gen5 keeps the eight GPUs fed; dual 25 GbE on-board with optional 100 GbE or InfiniBand for scale-out clusters. No PCIe contention, no storage starvation under sustained load.
Operates inside Malaysian DC reality
Validated for 200–240 V AC on dual hot-swap 2 kW Titanium PSUs, redundant cooling, ASPEED BMC for remote management — runs reliably in Malaysian DC environments without bespoke power conditioning.
Configurable without a redesign
GPU count (2 / 4 / 8), memory (256 GB – 2 TB), storage (8 – 60 TB), networking (25 / 100 GbE / InfiniBand), and CPU choice (AMD EPYC or Intel Xeon) — all selectable at quote without changing the chassis or rebuilding the BOM.
Architecture
Under the hood.
The four sub-systems that determine real-workload behaviour. We tune each before delivery.
GPU complex
- Up to 8 × NVIDIA RTX PRO 6000 Blackwell Server Edition (96 GB GDDR7 each)
- 24,064 CUDA cores · 752 5th-gen Tensor cores · 188 4th-gen RT cores per GPU
- FP4 / FP6 / FP8 inference acceleration (5th-gen Tensor)
- PCIe Gen5 ×16 per card — no NVLink on this generation of RTX PRO
CPU + memory
- Dual AMD EPYC 9745 (128 cores Zen 5 'Turin') or 9755
- Alternative: Dual Intel Xeon Platinum 8570 / Granite Rapids
- Up to 256 PCIe Gen5 lanes total across both sockets
- Up to 2 TB ECC DDR5-6400 across 24 DIMM slots (RDIMM)
Storage & networking
- Up to 60 TB U.2 / U.3 NVMe (PCIe Gen5)
- RAID 0/1/10 across NVMe; optional hot-swap SAS/SATA bays for archive
- Dual 25 GbE on-board (SFP28 or RJ45)
- Optional 100 GbE Mellanox / Broadcom NICs · InfiniBand HDR for cluster scale-out
Power, cooling, management
- Redundant 2 × 2000 W 80+ Titanium PSUs (200–240 V AC)
- Hot-swap fans across the chassis
- ASPEED BMC with Redfish API, iKVM, signed firmware updates
- Manufacturer factory QA per Giga Computing standard
Next step
Tell us your workload. EMARQUE sizes the AI Server and sends a quote.
Supported workloads
Reference workload categories.
Workload categories documented in the manufacturer's reference materials. Sizing is confirmed with your technical team during scoping.
Departmental RAG
Internal knowledge chat for 500 users
Run a private RAG stack with a 70B-class model against your document store, code repos, and ticketing system. Eight GPUs split four ways gives four logical inference endpoints sized for 100+ concurrent users each — the right shape for finance, legal, engineering, and ops to share one server.
Production inference
24/7 multi-tenant model serving
TGI / vLLM / Triton serving multiple fine-tuned variants of a base model. PCIe Gen5 isolation per card means no NVLink coherence overhead — the right architecture when each request fits on one GPU and you want predictable per-tenant throughput rather than tightly coupled training.
Vision + voice
Real-time multi-stream processing
Eight independent cards parallelise across camera or audio streams cleanly — object detection on 50 4K streams, speech-to-text on 200 concurrent calls, or pose estimation across a factory floor. Each stream gets a dedicated GPU slice with consistent latency.
Mixed workload
Inference today, fine-tune at night
Run production inference on six cards during business hours, reallocate to LoRA / QLoRA fine-tuning runs on all eight overnight. The BMC + Redfish API makes the rebalance scriptable; no separate dev cluster required.
Full spec sheet
Every line documented at quotation.
Configurable. Final BOM, GPU mix, RAM and storage, and networking topology are confirmed in writing at quotation.
- Reference platform
- Giga Computing G493-class (4U PCIe) — Supermicro / ASUS equivalents on request
- GPU
- Up to 8 × NVIDIA RTX PRO 6000 Blackwell Server Edition · 96 GB GDDR7 with ECC
- GPU memory bandwidth
- 1.6 TB/s per card
- GPU TDP
- 300 W passive (server-optimised) — 600 W active variant available
- GPU interconnect
- PCIe Gen5 ×16 per card (no NVLink — independent cards)
- Total GPU memory pool
- Up to 768 GB across 8 cards
- CPU
- Dual AMD EPYC 9745 (128 cores) or Intel Xeon Platinum 8570
- Memory
- Up to 2 TB ECC DDR5-6400 (24 channels, dual-socket)
- Primary storage
- Up to 60 TB U.2 / U.3 NVMe (PCIe Gen5, RAID 0/1/10)
- Bulk storage
- Up to 200 TB enterprise HDD (RAID Z2) — optional
- Networking
- Dual 25 GbE on-board (SFP28) · 100 GbE optional
- Power
- Redundant 2 × 2000 W 80+ Titanium PSUs, hot-swappable (200–240 V AC)
- Form factor
- 4U rackmount, 19" rails included
- Cooling
- Air-cooled with redundant hot-swap fans
- Management
- ASPEED BMC with Redfish API, iKVM, signed firmware
- OS
- Ubuntu Server 24.04, RHEL 9, Windows Server 2025
- Warranty
- Manufacturer warranty (Giga Computing / OEM) · NVIDIA GPU warranty
FAQ
Common questions about AI Server
What is the NVIDIA RTX PRO 6000 Blackwell Server Edition?
Server-optimised variant of the RTX PRO 6000 Blackwell — same Blackwell silicon and 96 GB GDDR7 memory, with passive cooling (300 W TDP) designed to be cooled by the server chassis airflow. Targets enterprise rackmount deployments where the active 600 W variant would be impractical.
What reference platform does EMARQUE use?
EMARQUE supplies the AI Server on the Giga Computing G493-class 4U PCIe reference platform validated for the RTX PRO 6000 SE thermal envelope. Equivalent reference designs from Supermicro or ASUS are available on request. Final BOM is documented at quotation.
Is this configuration customisable?
Yes. GPU count (2 / 4 / 8 cards), memory (256 GB – 2 TB), storage (8 – 60 TB NVMe), networking (25 GbE / 100 GbE / InfiniBand), and OS are all configurable. CPU choice (AMD EPYC vs Intel Xeon) is customer-selectable within the manufacturer's published configuration matrix.
What's the warranty?
Manufacturer warranty applies — Giga Computing / OEM warranty on the chassis and components; NVIDIA's warranty entitlement on the RTX PRO 6000 SE GPUs. EMARQUE handles local warranty claim processing through the manufacturer's authorised channel.
What support does EMARQUE provide locally?
EMARQUE handles in-country delivery, customs, commissioning, acceptance testing, and Tier-1 support response in Malaysia. Tier-2/3 escalation routes to the manufacturer per the warranty entitlement. Optional service contracts for extended response SLAs and on-site engineering visits can be added.
If I need NVLink-coherent multi-GPU or HGX-class, what's the upgrade path?
Step into NVIDIA DGX B200 or NVIDIA DGX B300 (or HGX B200 / B300 OEM platforms from Dell, Giga Computing, or Supermicro). Those configurations are presented on the respective NVIDIA DGX product pages — each shows both NVIDIA-branded DGX systems and HGX OEM alternatives.
Request configuration & quotation.
Manufacturer specifications and warranty terms apply. EMARQUE issues a formal quotation through your Key Account Manager.
02Talk to EMARQUE
Tell us about your workload.
Model size, concurrency, latency budget, deployment site. EMARQUE returns a quote in MYR within one Malaysian business day, sized to the workload — not the salesperson’s quota.
- 01
Key Account Manager
+6012 627 2280 - 02
Request for Quotation
business@emarque.co