SYNAPSE TOOLKIT — EDGE SILICON
EDGE AI · SILICON-PROVEN · ZERO CLOUD

Small models. Powerful edge.
AI computing, optimized.

Deploy efficient, compact AI models directly on the SYNAPSE processor. Fast, secure, and energy-aware computing for the next generation of intelligent devices.

Start Building — Explore NPU View Documentation
+ + + +
COMPUTE EFFICIENCY
LATENCY:
0.0ms
POWER:
0.0W
MODELS:
0 Active
REDUCING MODEL
FOOTPRINT BY
0%
SYNAPSE NPU chip — neural network on silicon
MODEL RUNNING:
OBJECT DETECTION
3.3M PARAMS · INT8● LIVE
7.2 MS LATENCY◆0.9 W TDP◆12 TOPS INT8◆3.3M PARAMS◆68 GB/S BANDWIDTH◆EDGE-NATIVE◆ZERO CLOUD◆OTA UPDATES◆AES-256 SECURE◆
01 / MODEL ZOOSYNAPSE NPU

Featured Small Models

VIEW ALL MODELS
VisionFlow v2
VF-2 · VISION PIPELINE
SIZE
3.3M
1.4 MB INT8
ACCURACY
91.8%
+0.6 vs v1
TASKS
2
detect · seg
MODEL CARD
AudioSense
AS-1 · AUDIO PIPELINE
SIZE
5.2M
2.1 MB INT8
ACCURACY
94.1%
12 kHz path
TASKS
2
asr · kws
MODEL CARD
EdgeGuard
EG-3 · SECURITY
SIZE
1.1M
0.5 MB INT8
ACCURACY
97.3%
0.4 W idle
TASKS
1
anomaly
MODEL CARD

※ ALL MODELS SHIP PRE-QUANTIZED TO INT8 AND ARE TUNED PER-SILICON FOR NPU0 / NPU1 TARGETS.

02 / PLATFORMFROM FRONTIER TO FORGE

From frontier to forge in four stages.

01

Distill

Frontier teachers compress task knowledge into 1–7M parameter students. Quality tracked per-layer, automatically.

02

Quantize

INT8 / INT4 with calibration on-device data. Footprint drops ~78% with <0.5% accuracy loss, guaranteed.

03

Compile

Graph compiler fuses ops for the NPU's 4,096 MAC array. Memory-mapped weights, zero-copy I/O, no runtime bloat.

04

Deploy

Signed OTA rollout to fleets with automatic rollback. Telemetry and training data never leave the device.

$ TOOLCHAIN: synapse-cli forge-rt npu-driver 2.1 ota-agent eval-harness
Wearables
0.4 W BUDGET
Drones
12 FPS DETECT
Medical
HIPAA-READY
Automotive
ASIL-B
Smart Home
<1 W IDLE
Industrial
−40 °C RATED
03 / SILICONSN-2049

The datasheet doesn't lie.

A 4 nm neural processor built for one job: run small models at absurd efficiency. No fans, no cloud round-trips, no compromise.

  • Deterministic latency — no thermal throttling surprises
  • Secure enclave with per-model key sealing
  • Runs from coin-cell to industrial rail supplies
DOWNLOAD FULL DATASHEET (PDF)
SILICON PROVEN · REV 2.1
SYNAPSE NPU — TECHNICAL DATASHEET
SN-2049 Neural Processor
FIG. 1 — Die micrograph
PROCESS4 NM FINFET
MACS / CYCLE4,096
PERF12 TOPS INT8
MEMORY8 GB LPDDR5
BANDWIDTH68 GB/S
TDP0.9 W TYP
PACKAGE12 × 12 MM BGA
SECURITYAES-256 · ENCLAVE
OPERATING−40…105 °C
INTERFACEPCIe 4.0 · SPI
SN-2049-0017 · FAB 42 · LOT 7C
FIG. 1 — DIE MICROGRAPH
SYNAPSE NPU REV 2.1
04 / BENCHMARKSMEASURED, NOT MARKETED

Numbers we'll sign our name under.

VISIONFLOW V2 · BATCH 1 · INT8 · AMBIENT 25 °C. FULL METHODOLOGY IN DOCS.

EFFICIENCYTOKENS / WATT
SYNAPSE NPU48
MOBILE GPU9
LAPTOP CPU2
SPEEDP50 FIRST TOKEN*
SYNAPSE NPU7.2 ms
MOBILE GPU41 ms
LAPTOP CPU210 ms
FOOTPRINTMODEL ON DISK
SYNAPSE INT81.4 MB
FP16 SMALL14 MB
FP32 SMALL90 MB

* HIGHER BAR = BETTER (INVERSE SCALE). CLOUD RTT EXCLUDED — THERE IS NO CLOUD.

05 / DEPLOYONE COMMAND

Ship to silicon in one command.

The Synapse CLI handles quantization, compilation and signed rollout. You bring the model; the forge does the rest.

  • ✓Auto-quantizes to INT8 with calibration report
  • ✓Targets npu0 / npu1 with per-rev kernel selection
  • ✓Signed OTA with automatic fleet rollback
  • ✓Telemetry stays on-device, always
forge — zsh · npu1 attached ● 0.9 W
We replaced a cloud vision pipeline with a 3.3M-parameter model on Synapse silicon. It paid for itself in eleven days.
ELENA MARSH — VP ENGINEERING, HELIX ROBOTICS
Helix Robotics Meridian Bio Orbital Systems Kestrel AI Northwind
06 / GET STARTED

Put intelligence
on every edge.

Free developer kit, three model credits, and a datasheet worth printing. No cloud required — that's the point.

Get Started — Free Talk to an Engineer
NO CREDIT CARD · NO CLOUD · NO 40-PAGE MSA