Start a build

Platform In detail

Silicon, framework, devices — the specifications.

Vishwakarma — four classes, one 4-topology NPU.

Specification V-3 W-1 E-7 S-9
Form factor PhonesSmart wearablesEdge AIServers
Compute 22 TOPS1.9 TOPS480 TOPS3.2 PFLOPS
Power / memory 6.4 W0.8 W210 W256 GB HBM
Highlight 3nm SoCSub-1WRack-scale edgeTrain + serve
Ships in Satchit Edge Phone—Satchit Edge Industrial 5.0Satchit Edge Server
Description A 3nm phone SoC fusing a 4-topology NPU with always-on vision. Runs Sudarshana on-device, no cloud round-trip.A sub-1W microchip for rings and watches. Streams sensor inference with continuous on-body learning.A rack-scale edge node for factories, retail and towers. Deterministic latency under heavy real-time load.A full training-and-serving accelerator for sovereign data centres. 256GB HBM, in-fab interconnect.

Satchit Edge — one device, three shapes, one common cloud.

  • S-9 class

    Server

    S-9 class rack unit for training and serving at datacentre scale.

    ↳ Connected to Sudarshana Cloud

  • V-3 class

    Phone

    V-3 class handset with always-on on-device vision.

    ↳ Connected to Sudarshana Cloud

  • E-7 class

    Industrial 5.0

    E-7 class controller for factory floors, retail and towers.

    ↳ Connected to Sudarshana Cloud

Sudarshana ML — compiled for the Vishwakarma, hallucination-free by design.

  • Graph compiler

    vw-graph

    Lowers your model into a closed set of verified Vishwakarma kernels.

  • Sparse runtime

    vw-sparse

    Static, deterministic scheduling across the 4-topology NPU.

  • On-device quantizer

    vw-quant

    Quantizes models on the device.

  • Fleet telemetry

    vw-signal

    Telemetry across the deployed fleet.

  1. 01

    Compile

    Your model is lowered by vw-graph into a closed set of verified Vishwakarma kernels. No interpreter, no escape hatches.

  2. 02

    Schedule

    The sparse runtime places work across the 4-topology NPU with static, deterministic scheduling — latency is a property of the graph, not the queue.

  3. 03

    Verify

    Every inference is checked against a formal bound computed at compile time. Precision drift and out-of-distribution inputs are caught before the answer leaves the chip.

  4. 04

    Return or refuse

    If the bound holds, the answer is returned with its proof attached. If it does not, the system says so — it never guesses.

04 Modes of usage

Two ways to run.

Mode 01 · Your silicon

Edge AI in your datacenter

Own the Vishwakarma hardware. We co-design the edge node with you, and the full Sudarshana stack — compiler, runtime, quantizer — runs entirely inside your walls. Your data never leaves.

Buy the hardware · deterministic latency

Mode 02 · Our silicon

Sudarshana Cloud — pay per inference

Skip the hardware. Send your workload to Sudarshana Cloud and pay only for the inferences you run — the same hallucination-free framework, billed as you use it.

Per-inference billing · zero capex

05 Reference deployments

Deployment architectures that ship.

  1. Case 01 · Mode 01 · Your hardware

    Industrial 5.0 factory floor

    An E-7 class Satchit Edge controller sits on the factory floor, running deterministic vision inspection and line-control inference under 10 ms — no cloud round-trip in the critical path. Fleet-wide model updates sync from Sudarshana Cloud between shifts.

    <10 ms

    line-control inference

    1. Line sensors feed the controller on the floor
    2. Satchit Edge · Industrial 5.0 · E-7 inference < 10 ms → PLC line controls
    3. Sudarshana Cloud model updates, synced between shifts

  2. Case 02 · Mode 01 · Your datacenter

    Sovereign data centre

    A rack of S-9 accelerators trains and serves inside your own walls. Nothing leaves the building: telemetry to Sudarshana Cloud is read-only and optional, and the in-fab interconnect keeps multi-node training on one fabric.

    0

    bytes of data egress

    1. S-9 × N training + serving, in-fab interconnect
    2. Inside your walls zero data egress · all models and data stay inside
    3. Sudarshana Cloud read-only telemetry (optional)

  3. Case 03 · Hybrid · Mode 01 + Mode 02

    Connected phone fleet

    V-3 class handsets run always-on vision and language inference entirely on-device. Heavy batch workloads spill over to Sudarshana Cloud, billed per inference, and model updates flow back down to the fleet bidirectionally.

    On-device

    vision + language, always on

    1. V-3 handsets always-on vision, on-device
    2. Batch workloads spill over, pay per inference
    3. Sudarshana Cloud bidirectional model-update sync

  4. Case 04 · Mode 02 · Pay per inference

    Retail wearable mesh

    W-1 rings and badges run on-body sensor inference at sub-1W. An in-store E-7 gateway aggregates the mesh, and anything beyond local scope is handled by Sudarshana Cloud with per-inference billing — no hardware capex for the retailer.

    Zero

    hardware capex for the retailer

    1. W-1 rings & badges on-body inference, sub-1W
    2. In-store gateway · E-7 aggregates the mesh
    3. Sudarshana Cloud beyond local scope, per inference

None of these match your shape? We co-design the deployment with you — tell us the workload and the walls it has to stay inside.

Co-design your next AI device.

Start a build