Feed multi-GPU training clusters with terabytes-per-second of dataset throughput. Extend host memory with a hybrid tier. Watch every joule from a single console.
A single Datafabrix DF-R1 rack, delivered as one purchase order, installed as one unit, supported as one system.
The specific metrics our customers ask about first — and where this rack is engineered to win.
Merchant silicon is called out explicitly. Datafabrix engineering is called out explicitly. No hand-wave categories.
| Layer | Component | Notes |
|---|---|---|
| Compute | 8 × GPU host servers · Intel Xeon Scalable · dual-socket | GPU hosts + Datafabrix DF-X Xeon cards for control plane |
| Fabric | 2 × Datafabrix PCIe Gen4 Fabric Backplane (DF-B4) | Tier-1 US PCIe Gen4 switch silicon inside a Datafabrix backplane |
| AI storage | AI Storage Expansion (DF-S01) · 24 × NVMe | Native PCIe endpoints across the fabric — GPUDirect Storage ready |
| GPU storage | GPU Storage Infrastructure (DF-S06) | Direct GPU-to-NVMe path — bypasses CPU I/O staging |
| Memory tier | Hybrid Memory Expansion (DF-S04) · CXL-ready | Extends host DRAM with fabric-attached memory |
| DCIM | All 8 DCIM modules pre-instrumented | Guardian · Thermal · StorageOS · Insight · Twin · Cloud · Vision · Conventional DCIM |
| Networking | Front-end 400 GbE / IB (customer-selected) | InfiniBand HDR/NDR or 400 GbE — customer chooses at order time |
| Cooling | Rear-door heat exchanger ready · air-cooled default | Liquid-cooled variant available for >40 kW density |
A GPU that spends 40% of its time waiting on data is a GPU you paid twice for. This rack was engineered to make that ratio 5% instead of 40% — with a PCIe fabric-attached storage tier that streams datasets at bus-native throughput.
Large models spill working sets past socket-DIMM limits. The hybrid memory tier lets your training framework use fabric-attached memory as a native NUMA node — training runs that used to swap or crash just run.
AI training racks run hot, draw a lot of power, and fail in expensive ways. Ours ships with eight DCIM modules already talking to every device — you know your PUE, your hot spots, and your predicted-failure windows from the first power-on.
When a training job crashes at 3 a.m. and you cannot tell if it was the GPU, the NVMe, or the fabric — there is one phone number, one support contract, one team accountable. Not eight.
Pre-train 7B–70B parameter models on your own data. Fabric-attached dataset tier keeps GPUs at >90% utilisation.
Sustain millions of tiny reads across image classification, segmentation, and detection workloads.
Iterate on production models with a hot dataset tier and a warm checkpoint tier on the same fabric.
Text + image + audio + video workloads pushed through a single fabric — no protocol translations.
Our delivery discipline is the single biggest reason customers pick us over DIY assembly or a locked AI system. Here is what the timeline actually looks like.
One call to align on workload, site, and timeline. Written quote with BOM, power/cooling profile, and delivery date inside one business week.
Rack assembled and validated at our integration facility. Firmware loaded. DCIM modules provisioned. Burn-in tested against the specific workload class.
Rack ships fully assembled. On-site installation by our team or a certified partner. Cabled, powered, network-integrated, DCIM streaming to your operators.
Workload deployed on the rack. Acceptance tests. Handover to your operations team, with our engineering on standby for the first 90 days.
Each Datafabrix rack is a curated combination of solutions. Below are the solutions that ship inside this rack. Click any card for the block diagram, connection topology, and per-subsystem specs.
Tell us your workload, your site, and your timeline. We will respond within one business day.