GIGABYTE HGX B300 G894 NVIDIA AI Server (AMD / Intel)

The GIGABYTE G894 is GIGABYTE's NVIDIA HGX B300 frontier-AI server — an 8U dual-processor system with 8× NVIDIA HGX B300 Blackwell Ultra GPUs on the HGX baseboard with full NVLink and NVLink Switch interconnect. EMARQUE offers it in two CPU configurations on one platform:

• G894-ZD3-AAX7 — dual AMD EPYC 9005 / 9004 (SP5, up to 192 cores per socket).
• G894-AD3-AAX7 — dual Intel Xeon 6900E+ / 6900-series (up to 128 P-cores or 288 E-cores per socket).

Each ships as a complete system: the 8U chassis with 8× HGX B300 GPUs, 2.1 TB HBM3E, onboard NVIDIA ConnectX-8 SuperNICs, 12-channel DDR5, Gen5 NVMe storage, 12× 3000W Titanium redundant PSUs, and GIGABYTE's management stack. EMARQUE AI Solutions are available on request — deployment scoping, rack and 3-phase PDU sizing, InfiniBand / Ethernet fabric design, NCCL and MPI tuning, multi-server cluster setup, orchestration, private AI stack deployment, team training. Final configuration confirmed on enquiry.

The GIGABYTE G894 follows GIGABYTE's published specification for this NVIDIA HGX B300 platform:

• GPUs: 8× NVIDIA HGX B300 Blackwell Ultra, full NVLink + NVLink Switch interconnect, 1.8 TB/s GPU-to-GPU bandwidth.
• GPU memory: 2.1 TB HBM3E across the 8-GPU baseboard.
• AI compute: up to 7× the AI compute of Hopper-generation platforms (NVIDIA Blackwell Ultra), built for AI reasoning and the most complex inference and training workloads.
• CPU: dual AMD EPYC 9005 / 9004 (G894-ZD3, up to 192 cores/socket) or dual Intel Xeon 6900E+ / 6900-series (G894-AD3, up to 128 P-cores or 288 E-cores/socket).
• System memory: AMD config — 12-channel DDR5 RDIMM, 24× DIMMs (EPYC 9005 up to 6400 MT/s). Intel config — 8-channel per CPU, 32× DIMMs, DDR5 RDIMM/MRDIMM.
• GPU networking: 8× 800 Gb/s OSFP ports via onboard NVIDIA ConnectX-8 SuperNICs — InfiniBand XDR or dual 400 Gb/s Ethernet. NVIDIA BlueField-3 DPU compatible.
• Storage: 8× 2.5-inch Gen5 NVMe hot-swap bays + 2× M.2 (PCIe Gen3).
• Expansion: 4× FHHL PCIe Gen5 x16 slots.
• Management LAN: 2× 10 Gb/s via Intel X710-AT2; GIGABYTE Management Console + GSM (IPMI / Redfish); optional TPM 2.0.
• Form factor: 8U rackmount, 19-inch standard, dual-processor.
• Power: 12× 3000W 80 PLUS Titanium redundant PSUs — requires datacentre-grade 3-phase rack power.
• Cooling: air-cooled with Automatic Fan Speed Control; liquid-cooled HGX B300 platforms available on enquiry.
• OS / Stack: NVIDIA AI Enterprise stack; DGX-OS-class environment scoped per deployment.

Exact configuration (CPU SKU, system memory, NVMe capacity, networking, cooling) and final pricing confirmed on enquiry — email business@emarque.co.

NVIDIA HGX B300 is NVIDIA's baseboard platform for frontier AI. The HGX baseboard mounts 8× B300 (Blackwell Ultra) GPUs with full NVLink and NVLink Switch interconnect (1.8 TB/s GPU-to-GPU), giving every GPU high-bandwidth access to every other GPU in the box. This is the same architectural pattern as DGX H100 / DGX H200 before it — one server, 8 GPUs, all-to-all NVLink — updated to Blackwell Ultra silicon with 2.1 TB HBM3E and up to 7× the AI compute of Hopper-generation platforms.

The GIGABYTE G894 is one of the HGX B300 server platforms shipping today. HGX B300 is the right platform for organisations training frontier LLMs on-prem, serving production inference at the highest throughput, building AI factories at multi-server scale, and meeting sovereign-AI requirements where data must stay in country. EMARQUE is among the first solution providers in Malaysia taking pre-orders for this tier.

The GIGABYTE G894 ships with GIGABYTE's 3-year manufacturer warranty covering hardware faults. EMARQUE's built-in role is Malaysia-based local handling:

• Single point of contact — pre-sale, post-sale and warranty routed through the same EMARQUE team.
• RMA routing — if a hardware fault appears, EMARQUE routes the GIGABYTE warranty claim on your behalf.

For deeper post-sale needs — pre-deployment burn-in, OS / driver / NCCL tuning, InfiniBand or Ethernet fabric design, multi-server cluster setup, orchestration, private AI stack deployment, team training — solutions are available from the EMARQUE AI team on request. Scope and charges quoted on enquiry. Software / OS / driver / setup support is scoped to the 3-year warranty window.

Step down vs the EMARQUE AI Server — 8× RTX PRO 6000 Blackwell: 8× RTX PRO 6000 is the production-density GDDR7-based tier — 768 GB aggregate GPU memory, PCIe Gen5 fabric, lower cost per box. the GIGABYTE G894 HGX B300 sits a tier above for frontier training and the highest-throughput inference: full NVLink GPU-to-GPU bandwidth, 2.1 TB HBM3E, designed to compose into multi-server training fabrics.

Step up vs NVIDIA Vera Rubin: Vera Rubin is the next-generation architecture after Blackwell Ultra. Choose HGX B300 if your deployment timeline is now and you need shipping-today hardware; Vera Rubin if your roadmap aligns with NVIDIA's launch timeline and the next-gen platform fits.

Above all of these sit rack-scale platforms (GB300 NVL72 and the Vera Rubin equivalent) — full NVLink Switch across an entire rack. EMARQUE AI scopes these on request.

Production AI workloads — the things that exceed a single workstation:

• Multi-tenant inference — serve multiple models or users concurrently with predictable latency. NIM microservices, vLLM, TensorRT-LLM.
• LLM fine-tuning at production scale — SFT, RLHF, DPO, LoRA, QLoRA on mid-to-frontier models.
• Retrieval-augmented generation at organisation scale — embedding pipelines, vector search, document AI for hundreds of thousands of documents.
• Computer vision and multimodal — VLM training and inference, video pipelines.
• Agentic AI with secure tool-calling and autonomous workflows.
• RAPIDS for massive-dataset data science.

Real-world: private GPT-class fine-tuning on enterprise data, organisation-wide internal LLM serving, sovereign-AI deployments where data must stay on-prem.

EMARQUE AI Servers ship with the full NVIDIA AI Enterprise stack — CUDA, container runtimes, NIM microservices for production inference, AI Workbench, and pre-built framework images (PyTorch, vLLM, RAPIDS, NeMo). NVIDIA HGX B300 additionally ships with NVIDIA DGX OS and Base Command for multi-tenant orchestration; EMARQUE-built RTX PRO 6000 servers ship on Ubuntu LTS plus the NVIDIA AI Enterprise stack.

EMARQUE AI Solutions on request can layer in a ready-to-use private LLM, RAG, embeddings and agentic-AI stack pre-configured for your data and use case — chat UI, vector search, document ingestion, model serving, all on-prem.

The GIGABYTE G894 HGX B300 platform is shipping now. Pre-ordering with EMARQUE:

• Locks your allocation for delivery in Malaysia.
• Final pricing confirmed on enquiry — pricing varies significantly by CPU configuration (AMD EPYC G894-ZD3 vs Intel Xeon G894-AD3), system memory, NVMe capacity, networking and cooling options, and EMARQUE AI services scope. Listed price is an indicative estimate.
• Delivery window aligned with GIGABYTE's Malaysia shipping schedule — typical lead time confirmed at order.
• Direct access to the EMARQUE AI team during pre-order for deployment scoping (rack, 3-phase PDU, InfiniBand or Ethernet fabric, multi-server cluster planning, software stack).

Click ENQUIRE NOW above or email business@emarque.co.

EMARQUE is among the first solution providers in Malaysia taking pre-orders for GIGABYTE G894 HGX B300 servers. What sets EMARQUE apart on this tier is the depth of solutions available from the EMARQUE AI team:

• Deployment scoping — rack and 3-phase PDU sizing, datacentre cooling specification (air vs direct-liquid), InfiniBand vs Ethernet fabric decision, multi-server cluster topology, NCCL and MPI tuning. Scope and charges quoted on enquiry.
• CPU configuration sizing — honest assessment of whether the AMD EPYC (G894-ZD3) or Intel Xeon (G894-AD3) configuration fits your workload, your rack and your support expectations.
• Orchestration — multi-tenant orchestration setup, model serving, internal AI platform layer.
• Local point of contact — pre-sale sizing, post-sale technical questions, warranty handling — same Malaysian team, no offshore queue.
• EMARQUE AI stack + assisted setup — pre-configured private stack (chat UI, RAG with vector search, document ingestion, model serving) installed and tuned for your data.

For buyers comparing HGX B300 providers, the practical question is who can integrate the server, the rack, the fabric and the AI stack — in country, with one point of contact. Email business@emarque.co for sizing.

Yes. EMARQUE AI is our dedicated AI compute division — full NVIDIA AI hardware stack and on-prem deployment for Malaysian businesses, research labs, and enterprises: NVIDIA Personal AI (DGX Spark, 2× Spark stacking), AI workstations (RTX PRO 6000 Blackwell and others), DGX Station GB300, multi-GPU AI servers, plus full integration support. Visit EMARQUE.AI or contact business@emarque.co.