Large-model serving
Build a dedicated inference service for larger models, with precision and throughput validated against your own requests.
All AI systems
Kulper AI System · DATACENTER AI
Frontier B300
Your own AI, at datacenter scale.
Loading rotation view…
Built around your workload Installed and ready to use
TALK TO KULPER
Let's plan your B300 system.
Talk to our team about your models, performance targets and installation site. We'll work through the configuration, current availability and support with you.
- GPU platform
- 8 × NVIDIA B300 · NVLink / NVSwitch
- Enclosure
- 8U rack mounted · air cooled · proposed design
- System memory
- 4 TB ECC RAM proposed · 2 TB to 8 TB options
- Storage
- Mirrored boot SSDs and 30.72 TB raw data SSDs proposed · larger capacities available
We confirm qualified components, rack space, power, cooling, networking and delivery as part of your project quote in $ USD. Ask us about purchasing, financing or leasing.
Hardware warranty & onsite service
Your quote specifies the selected hardware's warranty, onsite service options, response terms and regional coverage. Service levels are agreed for your installation, separately from optional model management.
Warranty & service detailsFrontier B300 · OVERVIEW
Dedicated infrastructure for demanding AI.
Frontier B300 combines an eight-GPU NVIDIA HGX platform with integration and support from Kulper AI Systems. We plan the build around your model, context length, concurrency and installation site, then test the agreed workloads before handover.
Training & fine-tuning
Scope model adaptation and training jobs around dataset size, GPU memory and the required software stack.
Shared AI infrastructure
Plan capacity, access and workload scheduling for multiple teams or business processes on your own infrastructure.
THE HARDWARE
Specifications & options.
- System
- Kulper AI System — Frontier B300 · proposed 8U air-cooled rack design
- GPU platform
- 8 × NVIDIA B300 SXM · HGX baseboard · NVLink / NVSwitch
- GPU memory
- 270 GB HBM3e per GPU · approximately 2.1 TB across eight GPUs, using NVIDIA's datasheet figures for planning
- Proposed CPUs
- 2 × Intel Xeon 6767P · 64 cores per CPU · 128 cores total · 2.4 GHz base
- ECC system RAM
- 4 TB default · 2 TB / 4 TB / 8 TB options · qualified, balanced DDR5 DIMM population
- Proposed boot drives
- 2 × 1.92 TB enterprise NVMe · mirrored · 1.92 TB usable before formatting
- Proposed data drives
- 4 × 7.68 TB enterprise NVMe (30.72 TB raw) by default; 8 × 7.68 TB (61.44 TB raw) option. Drive format and usable space depend on the final platform and agreed RAID layout.
- Proposed networking
- 8 × ConnectX-8, up to 800 Gb/s per adapter; dual 10 GbE connectivity. BlueField-3, fabric switches, optics and cabling scoped in your quote.
- Availability
- Contact Kulper AI Systems for pricing, availability and installation · USD, GBP or EUR
These are proposed Kulper AI System specifications. The final platform, qualified components, firmware and service package are confirmed in your quote. Kulper AI Systems handles configuration, software integration, acceptance testing and installation planning.
NVIDIA's HGX reference architecture lists up to 288 GB per GPU / 2.3 TB total. We use the lower datasheet figures above for planning and confirm available model memory on the supplied system.
Model fit still depends on precision, context and concurrency. NVLink supports communication between GPUs; software must distribute the workload. We validate your models before committing to performance targets.
Designed around your site.
- Rack & handling
- Plan for an 8U air-cooled system. Final chassis dimensions, weight, rack load, rails, airflow clearance and delivery access are confirmed for the selected build.
- Power & cooling
- Datacenter power and cooling are required. Supply voltage, circuits, PDU connections and cooling capacity are specified for the final system and its power limits. US / UK / EU identifies the installation region.
At 1,100 W maximum TDP per GPU, eight GPUs alone total 8.8 kW. CPUs, memory, networking and fans add to this. This is not a whole-server power or cooling estimate; the quote includes a site plan for the agreed configuration.
We check the completed system against the NVIDIA HGX reference design, including memory bandwidth and channel population, PCIe topology, networking and secure boot / TPM requirements.
NVIDIA HGX reference requirementsSSD figures are total raw capacity before formatting or RAID. Model fit and performance depend on quantization, context length, and concurrent use.
FROM HARDWARE TO WORKING AI
We help you get running.
Delivered ready. Installed your way.
We assemble and test the hardware, preload your chosen models and confirm your delivery date. Choose UPS shipping where suitable, or discuss delivery for a larger rack system. Install it yourself with our setup guide, or have our team connect the system, install clients and help assign user permissions. Shipping and installation work are itemized in your quote.
Managed by our team, or by you
Choose ongoing model updates, maintenance, and support, or a configured system handed over to your team. Management is quoted separately.
Explore Kulper ControlHardware warranty & onsite service
Hardware coverage follows the applicable component and platform manufacturer terms. Your quote identifies the supplied parts, coverage periods and how our team coordinates a repair or replacement.
Onsite service, labor, shipping, response times and extended coverage are agreed for your location before you order. Any service contract required by the selected platform is included in the proposal.
How hardware coverage worksBuy, finance, or lease
Discuss purchase, finance or lease options directly with our team. Terms and availability are confirmed for your project.