Bring the data to the computer — or the computer to the data.
NVIDIA builds two platform families on the same Blackwell GPU generation and the same CUDA ecosystem, engineered for opposite ends of the operational stack. DGX is data-centre infrastructure: it assumes conditioned power, cooling, and a network fabric oriented toward GPU-to-GPU communication, and it exists to train, fine-tune and serve large models against data brought to it[3]. IGX is industrial edge infrastructure: it assumes it is bolted into a machine or a cabinet, ingesting continuous multi-gigabit sensor streams and deciding within a bounded latency, supervised by an independent functional-safety domain, with a 10-year lifecycle[4][5]. Baru — the trusted intelligence-fusion platform — fuses where the mission runs; its silicon must follow the same rule.
Long description
Five platform tiles sit on one axis running from the edge to the data centre. Two principles label the halves. Under IGX — bring the AI computer to where data is generated — sit Jetson and IGX T5000 for embedded integration in a 40 to 130 watt envelope, and the IGX Thor T7000 as the sensor AI node, marked as the fusion hub of the fabric. Under DGX — bring the data to the AI computer — sit DGX Spark as a developer desktop, DGX Station GB300 deskside, and DGX B300 and NVL72 at rack and cluster scale. Moving left to right, model size increases and proximity to the sensor decreases.
Zero-copy, sensor to GPU.
High-rate sensor processing is governed less by peak TFLOPS than by how sensor bytes reach GPU memory. A conventional design routes every packet through the NIC, the kernel network stack, CPU buffers, system memory and a PCIe copy — each stage adding copies, CPU load, latency, and latency jitter. The IGX Thor T7000 integrates a ConnectX-7 SmartNIC on the board: sensor streams are written directly into GPU memory over RoCE using GPUDirect RDMA, with the transfer offloaded from the CPU, and the Holoscan SDK provides the pipeline from acquisition to inference[1][4][6]. This is the ingest architecture a fusion hub needs: dual 200 GbE ports give roughly 400 Gb/s of aggregate physical capacity — a five-sensor load of 100 Gb/s rides one port with headroom, and the second stays free for redundancy[1].
Long description
Two vertical flows, side by side. On the left, the conventional path in seven stages: sensor, NIC, kernel network stack, CPU buffers, system memory, a PCIe copy, then GPU memory and inference. Every stage adds copies, CPU load and jitter. On the right, the IGX path in six stages: high-speed sensors, the Holoscan Sensor Bridge, an on-board ConnectX-7 with two 200 gigabit Ethernet ports over RoCE, a direct write into GPU memory, Holoscan and TensorRT, then deterministic output. The direct write is the stage that removes the copies; the transfer is offloaded from the CPU by GPUDirect RDMA.
For real-time fusion the tail governs: 5 ms average with 12 ms worst-case supports the mission; 3 ms average with 80 ms at the 99.9th percentile does not.
Every tier, sized to its job.
Enlil — the Defence implementation of Baru — runs a four-tier fabric, each tier on NVIDIA silicon sized to its role[7]. The analysis behind this note is the engineering rationale for that mapping.
| Tier | Role | Silicon | Why |
|---|---|---|---|
| Tier 01 · FIB Node | Carried inference, peer-to-peer mesh | Jetson Orin / Thor-class SoM | 40–130 W envelope; embedded integration[5] |
| Tier 02 · FIB Hub | Cluster fusion centre; high-rate sensor aggregation | IGX Thor T7000 — the industrial, safety-certified build of the Thor SoM | ConnectX-7 dual 200 GbE ingest, GPUDirect RDMA, Functional Safety Island + Safety MCU (IEC 61508 / ISO 26262 design targets), BMC, 10-year lifecycle[1][4] |
| Tier 03 · FOB-IFC | Forward all-source correlation, ~80B-parameter models | Grace-Blackwell class (DGX Station GB300) | 748 GB coherent memory; deskside power envelope[8] |
| Tier 04 · Central IFC | Deepest all-source reasoning; training and fleet analytics | DGX B300 / NVL72 class | Data-centre training and large-model inference at scale[3][9] |
Jetson Thor and IGX Thor share the Thor system-on-module family; the IGX build adds the industrial hardware, the integrated SmartNIC, and the safety domain that production hubs require[4][5].
The functional-safety domain is an engineering property of the silicon; the line that governs action is Enkidu's own — it holds in every sector (see Trust).
DGX trains. IGX operates.
The platform decision is not DGX or IGX — it is both, in one loop. DGX-class infrastructure in the rear is the model factory: training, fine-tuning, synthetic data, fleet analytics, and the TensorRT optimisation that fits a model to its tier. DGX Spark on a developer's desk approximates the hub's sensor environment — the same ConnectX-7 family and the same Holoscan Sensor Bridge support — so pipelines are proven before they deploy[6][10]. IGX at the edge runs them against live streams within a bounded latency, and returns curated field data for the next cycle. This is the hardware half of Baru's model discipline — start strong, tune efficiently, fit the tier, shrink the hot path.
Long description
Three horizontal bands. At the top, the data centre as model factory: training, fine-tuning, synthetic data, fleet analytics and TensorRT optimisation. In the middle, development: pipelines proven against a representative sensor environment before deployment. At the bottom, marked as the operational tier, the physical edge: live streams handled within a bounded latency, returning curated field data. Optimised models and over-the-air updates flow downward through the bands; curated field data and telemetry flow back upward. Neither half stands alone — the rear trains what the edge runs, and the edge supplies what the rear trains on.
The dollar figures, in Canadian dollars.
Indicative only, for budgeting comparison: published USD prices and listings as of August 2026, converted at an assumed 1.37 CAD/USD, excluding taxes, duties, shipping and support. IGX hardware is quote-based through authorised distributors — in Canada typically Arrow, CDW Canada or Insight Canada — so rows marked “est.” require a formal quotation[11].
| Item | Approx. CAD | Note |
|---|---|---|
| Jetson AGX Thor Developer Kit | C$4,800 – C$7,550[12] | Closest low-cost proxy for Thor-class compute |
| DGX Spark | C$5,500 – C$6,450[10] | Per-developer workstation |
| IGX Orin Developer Kit | C$8,000 – C$14,000 (est.)[11] | Quote-based via distributors |
| RTX PRO 6000 Blackwell dGPU | C$11,200 – C$13,700[13] | Optional hub dGPU — 96 GB GDDR7 |
| IGX Thor kit + RTX PRO 6000 | C$18,000 – C$30,000 (est.) | Production-representative hub configuration |
| DGX Station GB300 | C$123,000 – C$158,000[8] | Tier 03 class |
| DGX B300 system | C$411,000 – C$480,000[9] | Tier 04 class; excludes fabric, storage, facility |
A complete development environment — one DGX Spark per developer and one shared, fully configured hub — assembles for roughly C$25,000–C$45,000, and rear-tier training capacity is better rented until sustained utilisation justifies owned hardware[9].
- [1]NVIDIA Technical Blog — “NVIDIA IGX Thor Powers Industrial, Medical, and Robotics Edge AI Applications,” April 2026.
- [2]NVIDIA — IGX platform overview (nvidia.com/en-eu/edge-computing/products/igx), accessed August 2026.
- [3]NVIDIA — DGX platform documentation, accessed August 2026.
- [4]NVIDIA IGX documentation — IGX Thor T7000 board kit.
- [5]NVIDIA — Jetson Thor / IGX Thor T5000 module documentation.
- [6]NVIDIA Holoscan SDK documentation — performance guidance and Sensor Bridge supported platforms.
- [7]Enkidu — Enlil conceptual architecture, v0.14, July 2026.
- [8]Tom’s Hardware; ServeTheHome; VRLA Tech — DGX Station GB300 listings and analyses, US$90,000–$115,000, 2026.
- [9]Tech-Insider — “NVIDIA Blackwell GPU Pricing,” August 2026: DGX B300 US$300,000–$350,000; B300 cloud US$7.10–$17.80/GPU-hr.
- [10]Engadget, October 2025; IntuitionLabs, 2026 — DGX Spark US$3,999–$4,699.
- [11]Arrow Electronics — IGX distribution; estimate basis in the whitepaper, Appendix A.
- [12]Hackster.io / CNX-Software, August 2025 — Jetson AGX Thor Developer Kit US$3,499; NVIDIA marketplace bundle US$5,499.
- [13]TechRadar / PNY — RTX PRO 6000 Blackwell US$8,200–$10,000.
Conversion basis and estimate construction: whitepaper, Appendix A. Canonical language of this note is English; translations are derived artifacts.
Read the full analysis.
The complete whitepaper — specifications, configurations, the weighted assessment, and the pricing annex — in four editions.