8× NVIDIA Blackwell Ultra GPUs
Eight NVIDIA B300 (Blackwell Ultra) GPUs on an NVIDIA HGX B300 baseboard. Fifth-generation NVLink and NVSwitch interconnect deliver a coherent 2,304 GB HBM3e memory pool — 60% more per GPU than NVIDIA DGX B200.
Pre-order / allocationNVIDIA DGX
NVIDIA DGX B300
Built for the era of reasoning AI — eight NVIDIA Blackwell Ultra GPUs in a unified, air-cooled DGX system.
Pricing on request — allocation and configuration confirmed at quotation.
8Blackwell Ultra GPUs
2,304 GBHBM3e
11×vs Hopper inference
Key features
Configuration overview.
Manufacturer-defined features from the published datasheet.
Up to 11× inference vs Hopper
Per NVIDIA's published figures, NVIDIA DGX B300 delivers approximately 4× faster training and up to 11× faster inference than the NVIDIA Hopper generation. Optimised for AI reasoning workloads.
8× NVIDIA ConnectX-8 SuperNICs
800 Gb/s per-GPU scale-out networking — twice the per-GPU bandwidth of NVIDIA DGX B200's ConnectX-7. Pairs with NVIDIA Quantum-X800 InfiniBand or NVIDIA Spectrum-X800 Ethernet rack fabrics.
NVIDIA Mission Control included
NVIDIA Mission Control software for cluster-level health monitoring, telemetry aggregation, validated recipe deployment, and job-scheduler integration. The standard operating model for NVIDIA DGX SuperPOD with B300.
NVIDIA AI Enterprise + DGX OS
Pre-installed software stack: NVIDIA DGX OS, NVIDIA AI Enterprise (latest), NVIDIA Base Command, NVIDIA NIM microservices, NVIDIA NeMo, and NVIDIA TensorRT-LLM optimised for Blackwell Ultra.
Three-year NVIDIA Enterprise Support
NVIDIA Business Standard Support hardware coverage and NVIDIA AI Enterprise software entitlement included as standard with every NVIDIA DGX B300 system.
Two ways to buy
Same platform — choose the supply path.
The NVIDIA HGX baseboard is identical on both paths. The turnkey DGX is fastest to deploy; OEM HGX platforms (Dell, Giga Computing, Supermicro) give wider configuration choice.
NVIDIA
NVIDIA DGX B300 (turnkey)
NVIDIA-built and NVIDIA-supported reference platform. Ships pre-configured with NVIDIA DGX OS, NVIDIA AI Enterprise (latest), NVIDIA Mission Control, NVIDIA Base Command, and a three-year NVIDIA Enterprise Support contract.
- 10U air-cooled NVIDIA DGX reference chassis
- NVIDIA DGX OS + NVIDIA AI Enterprise + NVIDIA Mission Control
- Three-year NVIDIA Enterprise Support contract included
- Reference building block for NVIDIA DGX SuperPOD with B300
Dell · Giga Computing · Supermicro
HGX B300 — OEM platforms
Same NVIDIA HGX B300 baseboard from NVIDIA's OEM partners. Customer-selectable CPU, memory, storage, networking, and cooling configurations within each OEM's published configuration matrix. Air-cooled and direct-liquid-cooled (DLC) variants available. OEM warranty and support model. NVIDIA AI Enterprise software available separately.
- Dell PowerEdge XE9712 (DLC) / XE9785 (HGX B300)
- Giga Computing G4L3 / G894-class (HGX B300, air or DLC)
- Supermicro SYS-A22GA-NBRT-LCC (HGX B300 liquid-cooled)
- Customer-selectable CPU, memory, NVMe topology, cooling
EMARQUE supplies both paths in Malaysia. Final configuration, lead time, and warranty terms are confirmed in writing at quotation.
Architecture
Under the hood.
The four sub-systems that determine real-workload behaviour. We tune each before delivery.
GPU complex
- 8 × NVIDIA B300 SXM (HGX B300 baseboard, Blackwell Ultra)
- 2,304 GB HBM3e total (288 GB × 8) — 60% more per GPU than B200
- 5th-gen NVLink + NVSwitch fabric · 1.8 TB/s per GPU all-to-all
- ~105 PFLOPS dense FP4 training · ~210 PFLOPS dense FP4 inference per system
CPU & system memory
- Dual Intel Xeon Platinum (8500-series, Emerald Rapids successor)
- Up to 4 TB DDR5
- PCIe Gen5 to HGX baseboard
- NVIDIA-validated CPU + memory configuration
Scale-out networking — ConnectX-8
- 8 × NVIDIA ConnectX-8 (per-GPU) — 800 Gb/s InfiniBand or Spectrum-X Ethernet
- 2× the per-GPU scale-out bandwidth of DGX B200's ConnectX-7
- Quantum-X800 / Spectrum-X800 rack-fabric compatibility
- GPUDirect RDMA across nodes
Power, cooling, software
- 10U rackmount, air-cooled — same form factor as DGX B200
- Power envelope sized for Blackwell Ultra density
- DGX OS · NVIDIA AI Enterprise (latest) · Mission Control · NIM · NeMo
- TensorRT-LLM optimised for B300 reasoning workloads
Next step
Get a DGX B300 configuration and lead time from your Malaysian NVIDIA systems specialist.
Supported workloads
Reference workload categories.
Workload categories documented in the manufacturer's reference materials. Sizing is confirmed with your technical team during scoping.
AI reasoning
Long-context inference and agentic workloads
NVIDIA positions DGX B300 for the AI reasoning era. 288 GB HBM3e per GPU supports the extended-context inference and multi-step agentic loops characteristic of reasoning models.
Generative AI
Foundation model training and fine-tuning
Per NVIDIA's published figures, approximately 4× faster training than NVIDIA Hopper generation. Coherent 2,304 GB HBM3e memory pool supports full-parameter fine-tuning of frontier-scale models in-node.
DGX SuperPOD
Reference building block for cluster scale-out
NVIDIA DGX B300 is the reference building block for NVIDIA DGX SuperPOD with B300 systems. Eight NVIDIA ConnectX-8 SuperNICs per system provide 800 Gb/s per-GPU scale-out via NVIDIA Quantum-X800 InfiniBand or Spectrum-X800 Ethernet fabrics.
In-place refresh
Same form factor as NVIDIA DGX B200
10U air-cooled chassis — same rack-U footprint as NVIDIA DGX B200. Existing DGX B200 site infrastructure (rack U, power circuits, cooling) typically carries over to NVIDIA DGX B300 deployments.
Full spec sheet
Every line documented at quotation.
As supplied by NVIDIA. EMARQUE handles in-country delivery, commissioning, and Tier-1 support handoff.
- GPU
- 8 × NVIDIA B300 (Blackwell Ultra) with NVLink 5
- GPU memory
- Up to 2.3 TB HBM3e per node (288 GB × 8)
- FP4 compute
- Up to 105 PFLOPS training · 210 PFLOPS inference per node (NVIDIA published)
- CPU
- Dual Intel Xeon Platinum (8500-series)
- System memory
- 4 TB DDR5
- Networking
- 8 × NVIDIA ConnectX-8 800 Gb/s InfiniBand
- Storage
- 30 TB internal NVMe (config varies)
- Form factor
- 10U rackmount, air-cooled
- Software
- NVIDIA DGX OS · AI Enterprise (latest) · NIM · NeMo
- Availability
- Pre-order / allocation — talk to EMARQUE for ETA
FAQ
Common questions about DGX B300
What is included with NVIDIA DGX B300?
Eight NVIDIA Blackwell Ultra (B300) GPUs on an NVIDIA HGX B300 baseboard with fifth-generation NVLink and NVSwitch, 2,304 GB total HBM3e (288 GB × 8), dual Intel Xeon Platinum CPUs (8500-series), 4 TB DDR5 system memory, internal NVMe storage, 8× NVIDIA ConnectX-8 SuperNICs at 800 Gb/s, NVIDIA DGX OS, NVIDIA AI Enterprise software, NVIDIA Mission Control, NVIDIA Base Command, and a three-year NVIDIA Enterprise Support contract. 10U air-cooled rackmount chassis.
What are the published performance figures?
Per NVIDIA: approximately 4× faster training and up to 11× faster inference compared to the NVIDIA Hopper generation. Exact petaFLOPS figures per NVIDIA's NVIDIA DGX B300 datasheet — confirm current numbers at quotation as NVIDIA continues to publish updated benchmarks.
How does NVIDIA DGX B300 differ from NVIDIA DGX B200?
Same 10U air-cooled form factor and same eight-GPU NVLink Switch domain. NVIDIA DGX B300 uses NVIDIA Blackwell Ultra (B300) GPUs with 288 GB HBM3e each (vs 180 GB in B200) — 60% more memory per GPU. Networking moves from NVIDIA ConnectX-7 (400 Gb/s) to NVIDIA ConnectX-8 (800 Gb/s). NVIDIA positions DGX B300 specifically for AI reasoning workloads.
How does NVIDIA DGX B300 differ from HGX B300 OEM systems?
Both use the same NVIDIA HGX B300 baseboard with eight Blackwell Ultra SXM GPUs. NVIDIA DGX B300 ships in NVIDIA's reference 10U chassis with NVIDIA DGX OS, NVIDIA AI Enterprise software, NVIDIA Mission Control, and a three-year NVIDIA Enterprise Support contract bundled. HGX B300 OEM systems from Dell (PowerEdge XE9712 / XE9785), Giga Computing (G4L3 / G894-class), and Supermicro (SYS-A22GA-NBRT-LCC) use the same baseboard but the OEM's chassis design and configuration matrix. Air-cooled and direct-liquid-cooled (DLC) chassis options are available from OEM partners.
Does NVIDIA DGX B300 require liquid cooling?
NVIDIA DGX B300 is air-cooled in the same 10U envelope as NVIDIA DGX B200. Direct-liquid-cooled HGX B300 systems are available separately from NVIDIA OEM partners (Dell PowerEdge XE9712, Supermicro SYS-A22GA-NBRT-LCC, and Giga Computing G4L3) for higher-density deployments where air cooling cannot extract the thermal load.
What is the typical lead time?
Lead time follows NVIDIA's allocation schedule for NVIDIA DGX B300. EMARQUE secures allocation through the NVIDIA channel and confirms projected delivery window at order acknowledgement. Allocation timing varies with order size and the current NVIDIA allocation window.
Request configuration & quotation.
Manufacturer specifications, factory lead times, and warranty terms apply. EMARQUE responds within one business day with a formal quotation and projected delivery window.
Contact Us
Get in Touch with Us
Tell us about your workload. We reply within one business day with a quote sized to fit.
- 01
Key Account Manager
+6012 627 2280 - 02
Request for Quotation
business@emarque.co