Cisco Architecture Proof of Concept

The Edge
Intelligence
Grid.

We transform Cisco Catalyst Switches and CW917X Access Points into a decentralized AI execution farm via Pipeline Parallelism. Zero dedicated GPUs. Total air-gapped sovereignty.

See the IOx Magic View Architecture

Beating the 100MB IOx limit.

Cisco CW917X APs have powerful ARM64 CPUs, but `cgroups` restrict memory to roughly 100MB. A 1B parameter model needs 2.2GB. Here is the math behind the magic.

01. Quantization

INT4 GGUF

We compress Llama-3.2-1B weights to 4 bits (Q4_K_M). Total model footprint plummets from 2.2GB to ~800MB with near-zero perplexity loss.

02. Pipeline Parallelism

Layer Slicing

The Llama-3.2-1B model contains 16 Transformer Layers. Instead of loading it fully, we slice the 800MB model horizontally across the network.

Total Size 800 MB

Total APs ÷ 16 Layers

Memory per Node ~50 MB

root@cat9300x-orchestrator:~

$ divaid-cli start --model llama3.2-1b --quant INT4

[INFO] Initializing Swarm Orchestrator...

[INFO] Discovering IOx Nodes on VLAN 100...

[OK] Found 16 active CW9178 Access Points.

[INFO] Sharding Llama-3.2-1B.gguf (812MB) into 16 parts...

[OK] Layer 0 deployed to 10.0.0.11 (48.1 MB)

[OK] Layer 1 deployed to 10.0.0.12 (48.1 MB)

... [Layers 2-14 deployed] ...

[OK] Layer 15 deployed to 10.0.0.26 (48.1 MB)

> Swarm Ready. Awaiting Prompts.

03. Execution

Success

50MB safely clears the Cisco 100MB hard limit. The container boots, the weights load into RAM, and the AP becomes a permanent Layer executor for the swarm.

Like a Blockchain for AI Operations.

Traditional inference requires a monolithic GPU holding the entire model in VRAM. We shatter this assumption.

Think of blockchain: millions of miners process tiny chunks of cryptographic work that chain together. Divaid uses your ceiling APs exactly like this.

Each AP computes its dedicated transformer layer and passes a microscopic 4KB hidden-state tensor to the next AP via Cisco's mGig backplane in microseconds.

Execution Pipeline

1. Orchestrator Tokenizer & Embeddings

2. AP Node 01 Attention + FFN (L0)

gRPC Tensor 4KB

3. AP Node 02 Attention + FFN (L1)

...