Armature Forge
248 TOPS
Armature Forge
248 TOPS
Heavyweight compute designed to scale
Compute
Draw
Thermal
Ingress
From
Compute
Draw
Thermal
Ingress
From
Compute
Draw
Thermal
Ingress
From
Approach
A sealed compute module that integrates silicon, firmware, and fleet tooling into one attested system reducing integration risk, cloud cost, and compliance exposure.
Approach
A sealed compute module that integrates silicon, firmware, and fleet tooling into one attested system reducing integration risk, cloud cost, and compliance exposure.
Integration
Same architecture and firmware image across the AR-9 line. Your code moves between Scout, Vector, and Forge without a rewrite.
Integration
Same architecture and firmware image across the AR-9 line. Your code moves between Scout, Vector, and Forge without a rewrite.
The Workhorse Module for Frontier Vision and On-Device Language Models
The AR-9 Vector is our workhorse module built for teams running frontier vision and on-device language models where the cloud isn't fast enough, allowed, or trusted. It ships sealed to IP67, runs fanless at a 14W typical draw, and holds deterministic latency across the full industrial thermal range.
The AR-9 Vector is our workhorse module built for teams running frontier vision and on-device language models where the cloud isn't fast enough, allowed, or trusted. It ships sealed to IP67, runs fanless at a 14W typical draw, and holds deterministic latency across the full industrial thermal range.
64GB LPDDR5X · 273GB/s Bandwidth
Level 3 Secure Enclave · Per-Unit Attestation
7-year contractual supply guarantee
The Workhorse Module for Frontier Vision and On-Device Language Models
The AR-9 Vector is our workhorse module built for teams running frontier vision and on-device language models where the cloud isn't fast enough, allowed, or trusted. It ships sealed to IP67, runs fanless at a 14W typical draw, and holds deterministic latency across the full industrial thermal range.
The AR-9 Vector is our workhorse module built for teams running frontier vision and on-device language models where the cloud isn't fast enough, allowed, or trusted. It ships sealed to IP67, runs fanless at a 14W typical draw, and holds deterministic latency across the full industrial thermal range.
64GB LPDDR5X · 273GB/s Bandwidth
Level 3 Secure Enclave · Per-Unit Attestation
7-year contractual supply guarantee
The Workhorse Module for Frontier Vision and On-Device Language Models
The AR-9 Vector is our workhorse module built for teams running frontier vision and on-device language models where the cloud isn't fast enough, allowed, or trusted. It ships sealed to IP67, runs fanless at a 14W typical draw, and holds deterministic latency across the full industrial thermal range.
The AR-9 Vector is our workhorse module built for teams running frontier vision and on-device language models where the cloud isn't fast enough, allowed, or trusted. It ships sealed to IP67, runs fanless at a 14W typical draw, and holds deterministic latency across the full industrial thermal range.
64GB LPDDR5X · 273GB/s Bandwidth
Level 3 Secure Enclave · Per-Unit Attestation
7-year contractual supply guarantee
Specs first.Adjectives never.
Verified 2026.06
Compute
Sustained throughput, INT8
Sustained throughput, INT8
Peak throughput, INT4
496 TOPS
Sustained throughput, INT8
4× on-die
Memory & storage
System memory
System memory
Peak throughput, INT4
496 TOPS
Sustained throughput, INT8
4× on-die
Power & thermal
System memory
System memory
Peak throughput, INT4
496 TOPS
Sustained throughput, INT8
4× on-die
I/O
System memory
System memory
Peak throughput, INT4
496 TOPS
Sustained throughput, INT8
4× on-die
Security & compliance
System memory
System memory
Peak throughput, INT4
496 TOPS
Sustained throughput, INT8
4× on-die
Specs first.Adjectives never.
Verified 2026.06
Compute
Sustained throughput, INT8
Sustained throughput, INT8
Peak throughput, INT4
496 TOPS
Sustained throughput, INT8
4× on-die
Memory & storage
System memory
System memory
Peak throughput, INT4
496 TOPS
Sustained throughput, INT8
4× on-die
Power & thermal
System memory
System memory
Peak throughput, INT4
496 TOPS
Sustained throughput, INT8
4× on-die
I/O
System memory
System memory
Peak throughput, INT4
496 TOPS
Sustained throughput, INT8
4× on-die
Security & compliance
System memory
System memory
Peak throughput, INT4
496 TOPS
Sustained throughput, INT8
4× on-die
Specs first.Adjectives never.
Verified 2026.06
Compute
Sustained throughput, INT8
Sustained throughput, INT8
Peak throughput, INT4
496 TOPS
Sustained throughput, INT8
4× on-die
Memory & storage
System memory
System memory
Peak throughput, INT4
496 TOPS
Sustained throughput, INT8
4× on-die
Power & thermal
System memory
System memory
Peak throughput, INT4
496 TOPS
Sustained throughput, INT8
4× on-die
I/O
System memory
System memory
Peak throughput, INT4
496 TOPS
Sustained throughput, INT8
4× on-die
Security & compliance
System memory
System memory
Peak throughput, INT4
496 TOPS
Sustained throughput, INT8
4× on-die
Every gram is purpose -built load-bearing
Four layers, one sealed unit. Nothing decorative, nothing you'll need to source separately.
A1
Anodized enclosure rated for washdown, dust, and 1m submersion.
A1
Anodized enclosure rated for washdown, dust, and 1m submersion.
A2
Fanless thermal mass tuned for 14W continuous at 65°C ambient.
A2
Fanless thermal mass tuned for 14W continuous at 65°C ambient.
A3
248 TOPS sustained, 64GB LPDDR5X, FIPS 140-3 secure enclave on-die.
A3
248 TOPS sustained, 64GB LPDDR5X, FIPS 140-3 secure enclave on-die.
A4
6× MIPI CSI-2, 2× 10G SFP+, CAN-FD, and 2× M.2 NVMe on one board.
A4
6× MIPI CSI-2, 2× 10G SFP+, CAN-FD, and 2× M.2 NVMe on one board.
Benchmarks
Measured, not marketed
Model
Latency
Throughput
Power
YOLOv9-S · 640×640
Object detection
2.8 ms
357 FPS
12.4 W
Llama-3.2 · 3B Q4
Object detection
18 ms/tok
55 tok/s
13.9 W
SAM 2 · base+
Segmentation
14 ms
2.8 ms
15.1 W
Model
Latency
Throughput
Power
YOLOv9-S · 640×640
Object detection
2.8 ms
357 FPS
12.4 W
Llama-3.2 · 3B Q4
Object detection
18 ms/tok
55 tok/s
13.9 W
SAM 2 · base+
Segmentation
14 ms
2.8 ms
15.1 W
Model
Latency
Throughput
Power
YOLOv9-S · 640×640
Object detection
2.8 ms
357 FPS
12.4 W
Llama-3.2 · 3B Q4
Object detection
18 ms/tok
55 tok/s
13.9 W
SAM 2 · base+
Segmentation
14 ms
2.8 ms
15.1 W
PRODUCT LINEUP
Related Armature lineup
Get started
An Eval Kit On Your Bench By Next Friday.
Fast shipping 2-3 days
No commitment try & Test
Full SDK access Documented
Full SDK access Documented
