Sign in Schedule a meeting

SENSORLOGIX · FORWARD DEPLOYMENT

What runs forward, and what it runs on

A collector beside each sensor, and a forward node where the analysis has to keep going with no link. This is what each one is built from, what it runs, and roughly what it carries.

The forward node runs the whole DataLogiX platform with SensorLogiX, under the same rules as the rear. It installs from media and never calls home.

PDF · 7 pages · hardware, software, models, capacity and standards

AT A GLANCEAPPROXIMATE

Three tiers, one set of rules

  • Edge collectorTIER 1Jetson Thor · about 220 W
  • Forward nodeTIER 21 GPU, 96 GB · about 1.2 kW
  • Forward node, largeTIER 32 nodes · about 2.5–3 kW

THE TIERS

Three tiers, sized to where they go

A vehicle or a mast gets a collector. A ship, a base or a command post gets a forward node, and a larger ship or a headquarters gets two.

TIER 1

Edge collector

WHERE

Vehicles, drone ground stations, masts and aircraft.

WHAT IT DOES

Captures and buffers sensor data through a lost link, with optional detection on the box. It doesn’t run analyses.

HARDWARE

A small sealed rugged box with an NVIDIA Jetson Thor module.

LANGUAGE MODEL

gpt-oss-20b, optional

POWER

About 220 W

TIER 2

Forward node

WHERE

Ships, forward bases and incident command posts.

WHAT IT DOES

The whole platform for one site, fully disconnected for days.

HARDWARE

One short-depth rugged 2U server with an NVIDIA RTX PRO 6000 Blackwell Max-Q, 96 GB.

LANGUAGE MODEL

gpt-oss-120b

POWER

About 1.0–1.2 kW

TIER 3

Forward node, large

WHERE

Larger ships and headquarters elements.

WHAT IT DOES

The whole platform with more video streams, longer retention and failover.

HARDWARE

Two rugged 2U servers, each with one 94–96 GB GPU: NVIDIA H100 NVL or RTX PRO 6000.

LANGUAGE MODEL

gpt-oss-120b

POWER

About 2.5–3 kW

  • Sync. Collectors upload to a forward node or the rear. Forward nodes sync both ways with the rear over a network, on removable media or through your cross-domain device.

HARDWARE

The hardware, tier by tier

The chassis and its GPU slots are the parts to check first: they decide what fits. The PDF lists a minimum and a recommended build for each tier.

EDGE COLLECTOR
FORWARD NODE
FORWARD NODE, LARGE
Form
Small sealed rugged box
Short-depth rugged 2U
Two rugged 2U servers
Compute
NVIDIA Jetson T5000 (Thor), 128 GB; T4000, 64 GB
32–48 x86-64 cores, 256–384 GB ECC
48–64 cores, 384–512 GB ECC per node
GPU
Built into the module
1 × RTX PRO 6000 Blackwell Max-Q, 96 GB
1 × H100 NVL, 94 GB, or RTX PRO 6000 per node
Storage
1–4 TB NVMe, self-encrypting
8–15 TB usable, mirrored NVMe, self-encrypting
30–60 TB usable, replicated between nodes
Network
1–2 × GbE, or 2.5/10 GbE
2 × 10/25 GbE and a management port
2 × 25 GbE per node, plus a node-to-node link
Keys
TPM 2.0; USB HSM optional
TPM 2.0 and a FIPS 140-3 Level 3 HSM
TPM 2.0 and a PCIe HSM per node
Power
18–32 VDC vehicle or 28 VDC aircraft; about 220 W
100–240 VAC or 28 VDC; 1.0–1.2 kW; UPS
About 2.5–3 kW; UPS for both

Why two nodes: one 94–96 GB GPU already runs gpt-oss-120b. The second node adds failover, keeps video work apart from the model and adds capacity.

See the GPUs we build withSee less
NVIDIA GPU
MEMORY
POWER
MIG PARTITIONS
USED FOR
RTX PRO 6000 Blackwell Max-Q
96 GB GDDR7
300 W
Up to 4
Forward node
H100 NVL
94 GB HBM3
350–400 W
Up to 7
Large forward node
L40S
48 GB GDDR6
350 W
None
Smaller builds, with gpt-oss-20b
L4
24 GB
72 W
None
Video only, or a small model
Jetson T5000 / T4000
128 / 64 GB
40–130 / 40–70 W
Yes
Edge collector

WHAT RUNS WHERE

What each tier does, and what waits

Forward, the checks are the cloud’s: the SQL guard, number bindings, recomputation, the verifier, traceability, markings and dissemination run exactly as they do in the rear. A capability that can’t meet its standard on local hardware is switched off, never run in a weaker form.

CAPABILITY
EDGE COLLECTOR
FORWARD NODE
LARGE FORWARD NODE
FORWARD NODE WITH NO GPU
Capture, buffering, store-and-forward
Yes
Yes
Yes
Yes
Ingest and parsing, entity search, maps, dashboards, review, dissemination, markings and traceability
No
Yes
Yes
Yes
New analyses and reports, written by the model
No
Yes
Yes
Queued, with a notice, until a GPU or a link is available
Video detection and tracking
Optional
Yes
Yes
Queued
OCR of scanned pages
No
Yes
Yes
Tesseract only, less accurate
Semantic and image search
No
Yes
Yes
Keyword and entity search only

Connected extras, such as Microsoft 365, Purview, Google Drive and hosted models, run on a forward node only when a link is up and policy allows them.

In FIPS mode, features whose algorithms sit outside FIPS 140-3 are off on every tier: post-quantum signatures, OpenPGP, C2PA content credentials and X25519/Ed25519.

SOFTWARE

The same software everywhere it runs

The forward node runs the same signed build as the rear. It needs no network to install, run or update.

LAYER
WHAT RUNS
Operating system
Red Hat Enterprise Linux 9, or Ubuntu 24.04 LTS with Ubuntu Pro, in FIPS mode from install. Disks encrypted with LUKS, the key sealed to TPM 2.0.
Kubernetes
RKE2 in FIPS mode, hardened to the DISA STIG and the CIS benchmark, with Pod Security “restricted” and audit logging always on. One node, or two for the large forward node.
Ingress
Traefik
GPU
NVIDIA driver, container toolkit and GPU operator, with MIG partitions.
Database
PostgreSQL 17 with pgvector and PostGIS
Object storage
SeaweedFS, S3-compatible
Workflows
Temporal, self-hosted
Keys and secrets
OpenBao, sealed by the HSM over PKCS#11 or by Shamir key shares. Envelope keys are rooted in the HSM.
Identity
Keycloak with CAC/PIV smart cards, OIDC and SAML, in FIPS mode. A local directory, mapped to the rear’s on the first sync.
Registry and updates
zot with cosign signatures and TUF-checked content. Bundles on media, delta updates and rollback. No call-home.
Monitoring
OpenTelemetry collector, Prometheus, Grafana and Loki
Backup
pgBackRest and Velero, to encrypted removable media
Time
chrony and gpsd, with PPS from the GPS receiver

Models

PURPOSE
EDGE COLLECTOR
FORWARD NODE
LARGE FORWARD NODE
Planner, writer and checker
gpt-oss-20b, optional
gpt-oss-120b (MXFP4, about 65 GB), or gpt-oss-20b
gpt-oss-120b
Second checker, from another model family
No
An open model of about 8–30B parameters, such as Qwen3 or Mistral Small
The same, larger
Embeddings
No
A local text embedder; SigLIP 2 for images
The same
Video detection and tracking
RF-DETR and ByteTrack on ONNX Runtime
The same, plus a captioner. Face detection is off unless policy turns it on
The same
OCR
No
PaddleOCR-VL on the GPU; Tesseract as fallback
The same
Speech to text
No
Local speech recognition
The same
Forecasting
No
Chronos-2 and statistical models, on the CPU
The same

Models ship inside the bundle and are checked by hash before they load. All are open-weight, under permissive licenses such as Apache-2.0. The checker is never from the writer’s model family.

See the DataLogiX services on the nodeSee less
  • Platform. The DataLogiX API, web app and workflow engine, webhooks and the relay gateway.
  • Documents. Conversion of Office, PDF and older formats; OCR; Word, PDF, PowerPoint and Excel reports; citation styles; XBRL and C-CDA readers; malware scanning with offline signature updates.
  • Security. OpenTDF key access, a local prompt screener that fails closed, and a code sandbox with no network.
  • Ingest. The relay, which on a collector captures video, splits out KLV metadata and stores and forwards; SFTP drop; watched folders; connectors.

APPROXIMATE CAPACITY

Roughly what each tier carries

All figures here are approximate and meant for sizing. Real numbers depend on the hardware, the questions, the data and how busy the site is, so we size each build with you.

FORWARD NODE
LARGE FORWARD NODE
Named users
25–50
75–150
People working at once
10–15
30–50
Analyses running at once
2–3
5–8
Question to verified report, typical
10–25 minutes
6–15 minutes
Entity and document search
Under 1 second, across about 10 million passages
The same, with more headroom

Search, maps, review and dashboards don’t use the GPU; the language model sets the pace. A report takes about 30,000–60,000 generated tokens across the planner, writer and checkers.

Video and tracks

EDGE COLLECTOR
FORWARD NODE
LARGE FORWARD NODE
Live 1080p streams with detection, at about 5 sampled frames a second each
1–2, on the box
4–6, while the model serves users
10–16
Recorded video
No
About 4–8× real time per GPU partition
About twice the forward node
Track and message updates on the live map (CoT, AIS, ADS-B)
Forwarded
Within about 5 seconds of receipt
The same

Detections arrive within about 2 minutes of each 30-second video segment closing.

See model speed, documents and storageSee less

Model speed

MODEL AND GPU
ONE REQUEST
MANY AT ONCE
gpt-oss-120b on H100 NVL
80–150 tokens/s
Several hundred tokens/s
gpt-oss-120b on RTX PRO 6000 Max-Q
50–110 tokens/s
200–400 tokens/s
gpt-oss-20b on RTX PRO 6000 or L40S
120–250 tokens/s
500+ tokens/s
gpt-oss-20b on Jetson T5000
20–50 tokens/s
No

Documents and data

FORWARD NODE
LARGE FORWARD NODE
Scanned pages through OCR
1,000–2,000 an hour, off-peak
2,500–5,000 an hour
Born-digital documents
3,000–6,000 pages an hour
6,000–12,000 pages an hour
Structured loads (CSV, database extracts)
10–50 million rows an hour
20–100 million rows an hour

Storage

EDGE COLLECTOR
FORWARD NODE, 8–15 TB
LARGE FORWARD NODE, 30–60 TB
Video
Hours to days of buffer; the oldest dropped only by policy
About 2–4 weeks of raw video from four 1080p streams, at about 2.7 TB a week, unless policy sends raw video to the rear and keeps clips
3–6 months at four streams, or more streams
Models and software
20–50 GB
150–250 GB
150–250 GB
Data, documents and indexes
No
Typically 1–3 TB
Typically 5–10 TB

Running it

  • Install. From media, with no network, in about 30 minutes.
  • Disconnected. 72 hours and more with no link at all.
  • Cold start. About 5–10 minutes from power to serving, plus the HSM or operator unseal.
  • Zeroize. Keys destroyed in seconds, leaving the data unreadable; then the disks are wiped.
  • Updates. Signed delta bundles on media, checked before they install, with rollback.
  • Sync with the rear. Resumable, by priority and within a bandwidth cap. Every record is checked against markings and release rules on both sides.

STANDARDS

Standards the kit is built to

We don’t certify hardware. The kit is specified to these standards, the hardware vendor’s certification of the chosen model qualifies it, and accreditation stays with you.

STANDARD
COVERS
TIERS
WHO HOLDS IT
MIL-STD-810H
Environment
All
Hardware vendor
MIL-STD-461H
Electromagnetic interference
All
Hardware vendor
MIL-STD-1275F
28 VDC vehicle power
1, 2
Hardware vendor
MIL-STD-704F
Aircraft power
1
Hardware vendor
MIL-DTL-901E, MIL-STD-167-1A
Shipboard shock and vibration
2, 3
Hardware vendor and shipyard
RTCA DO-160G
Airborne equipment
1
Hardware vendor
IEC 60945, IEC 60529, IEC 62368-1, FCC Part 15B, NEBS
Marine equipment, ingress protection, safety, emissions, telecom sites
As needed
Hardware vendor
FIPS 140-3
Cryptographic modules
All
The OS, drive and HSM vendors
DISA STIGs
Hardening of RKE2, Kubernetes and the OS
2, 3
We configure; your assessor checks
TEMPEST, CSfC, authority to operate
Emanations, classified networks, accreditation
2, 3
You
TAA, NDAA §889
Country of origin, covered telecommunications
All
Hardware vendor; EVAS by representation

SECURITY ON THE BOX

Locked down before it ships

The node keeps the rear’s controls with no link at all, and it’s built to be inspected.

  • Encrypted at rest. Self-encrypting drives under LUKS, per-tenant data keys wrapped by the HSM, and sealed columns encrypted just as in the cloud.
  • Smart-card sign-in. CAC/PIV, with no shared accounts.
  • Audited. Every action is recorded as it is in the cloud, and the audit trail syncs to the rear.
  • No outbound network. Egress is denied by default. Connected extras run only by policy, and every attempt to reach out is recorded.
  • Signed where it’s made. Reports and traces are signed with the node’s own key and timestamped from GPS time. The rear checks each signature on sync.
  • Evidence for your assessor. Software bills of materials, scan results, STIG checklist results and a control mapping come with every build.

Tell us where it has to run

We’ll size the kit with you, from the sensors and links you have to the people who will use it.

Book a briefingDownload the specification