Compute Routing Protocol // Agent Economy

Super Inu

For agents that need compute

Super Inu #001 · 1 of 1One mascot. One router.

Super Inu finds, scores and routes AI workloads to the best GPUs available, and explains every call.

Route a job Read the docs

Investment Case // Thesis

Own the decision, not the metal.

01 // Why now

Agents are about to become buyers

Software that launches its own training, inference and batch jobs makes the human choosing an instance type the bottleneck. Whoever serves that buyer first sets the standard.

02 // The wedge

Explainable placement, shipping today

Open GPU telemetry in, scored and justified recommendations out. v0.1 is built, tested and runs on mock data. Real execution is the next milestone.

03 // The moat

Every decision makes the next one better

Routing data compounds: which providers delivered, which GPUs held up, what actually ran fast. That history is what pricing, reputation and trust get built on.

04 // The prize

A market that routes itself

If compute becomes a programmable commodity, the router that sits between agents and every provider becomes the default interface to it.

05 // Neutral

Multi-vendor by design

One provider interface covers local clusters, clouds and marketplaces. No single supplier owns the customer.

06 // Discipline

Routing first, payments last

No token and no speculation in the core. Settlement arrives only after routing and verification are proven.

Today

Human picks a cloud
Human picks a GPU
Human provisions infrastructure
Workload runs
Human pays

With Super Inu

Agent states the job and its limits
Super Inu discovers and scores compute
It explains why it chose a target
Work is dispatched and verified
Compute is released, payment follows

Protocol Suite // Compute Router

Advanced compute routing modules

01 // Module

Telemetry Reader

Pulls GPU metrics from NVIDIA DCGM through dcgm-exporter and normalizes them into one GPU model.

GPU NodeDCGMdcgm-exporterNormalizer

02 // Module

Inventory

One live list of every GPU across providers: model, VRAM, utilization, temperature, power and health.

ProvidersInventoryReserved GPUs

03 // Module

Placement Engine

A deterministic 100-point score behind hard filters. Same inputs, same answer, no black box.

Hard FiltersFit 20Compute 25Memory 20Health 15Model 10Thermal 10

04 // Module

Explainer

Every recommendation says why: free VRAM, utilization, temperature, XID status, hardware match and warnings.

ReasonsWarningsRejections

05 // Module

Provider Interface

One contract (reserve, launch, status, release) so local clusters, clouds and marketplaces plug in alike.

LocalCloudMarketplaceComputeProvider

06 // Module

Execution Planner

Turns a recommendation into ordered steps. Plans only in v0.1; real execution is next.

ValidateSelectReserveLaunchMonitorVerifyRelease

Live product model // Streaming (simulated)

Watch the router think.

Agents submit jobs in real time. Super Inu scores the fleet, routes each job and streams the flow into the GPUs while telemetry drifts every second. The stream is simulated in your browser; the scoring matches the v0.1 code.

Arrival rate Job type GPUs Free VRAM Prefer

idlebusy· green ring = running a job · white flash = just routed · dashed red = unhealthy · click a GPU to inspect or break it

Select a GPU on the map to inspect it.

GPU tech

The signals behind every decision.

Super Inu reads open DCGM telemetry for H100, H200, A100, L40S, L4 and RTX-class GPUs wherever the metrics are exposed. Five signals drive v0.1. Four more unlock smarter routing next.

GPU utilization

How busy the compute engines are.

DCGM_FI_DEV_GPU_UTIL

v0.1 reads this

Framebuffer memory

Free VRAM decides whether a job fits.

DCGM_FI_DEV_FB_USED / FB_FREE

v0.1 reads this

Temperature

Thermal headroom and health.

DCGM_FI_DEV_GPU_TEMP

v0.1 reads this

Power draw

Enforces per-GPU power limits.

DCGM_FI_DEV_POWER_USAGE

v0.1 reads this

XID errors

Driver-reported faults mark a GPU degraded.

DCGM_FI_DEV_XID_ERRORS

v0.1 reads this

Tensor-core activity

Real AI-math load, not just 'busy'.

DCGM_FI_PROF_PIPE_TENSOR_ACTIVE

planned

NVLink bandwidth

Multi-GPU jobs need fast links.

DCGM_FI_PROF_NVLINK_TX_BYTES

planned

ECC errors

Memory reliability for long runs.

DCGM_FI_DEV_ECC_DBE_VOL_TOTAL

planned

Clocks and throttling

Spot GPUs slowing themselves down.

DCGM_FI_DEV_SM_CLOCK

planned

Field names come from NVIDIA DCGM; confirm them against the docs and NVIDIA's repos before relying on them.

Supported Providers

Route across clusters, clouds and marketplaces

Provider // 01

Local cluster

GPUs visible through a dcgm-exporter endpoint.

Live in v0.1

Provider // 02

Cloud A

First public-cloud provider behind the same interface.

Planned · v0.3

Provider // 03

Cloud B

A second cloud, so routing has real choices.

Planned · v0.3

Provider // 04

GPU marketplaces

Open markets of rentable compute, scored like any other source.

Planned · v0.3 to v0.6

Flywheel // Value Engine

The routing flywheel

Agent Job→Discover→Score→Route→Verify→Reputation Data→Better Routing ↻

There is no token in the core. Every run adds data about which providers delivered, so the next decision is better.

Leverage it for your next

Workload

Route a job

Coming Soon // Roadmap

The path from router to compute economy.

V0.1 · GPU intelligence
Telemetry, inventory, explainable placement, API. Built.

V0.2 · Workload execution
Real reserve, launch and release on Kubernetes with the NVIDIA GPU Operator.

V0.3 · Multi-provider compute
Cloud and marketplace providers behind one interface.

V0.4 · Pricing and optimization
Cheapest, fastest or safest objectives; budgets enforced.

V0.5 · Provider reputation
Reliability history and SLA tracking.

V0.6 · Autonomous marketplace
Agents negotiate directly with providers.

V1.0 · Agent-native compute economy
Verified execution and automatic settlement.

← Back to home

Documentation · v0.1

Super Inu docs

Super Inu turns GPU telemetry into explainable placement decisions. This page covers setup, the data it reads, the API and the NVIDIA open-source projects it builds on.

Quickstart

git clone <your-repo> && cd super-inu-compute
docker compose up                       # API on :8000, mock GPU data, no GPU needed
curl -X POST localhost:8000/recommend \
  -H 'content-type: application/json' -d @examples/workload.json

What it reads from NVIDIA DCGM

DCGM field	Super Inu field	Used for
`DCGM_FI_DEV_GPU_UTIL`	gpu_utilization	Compute headroom (25 pts)
`DCGM_FI_DEV_FB_USED` / `FB_FREE`	memory_gb, memory_used_gb	VRAM fit filter, memory headroom (20 pts)
`DCGM_FI_DEV_GPU_TEMP`	temperature_c	Thermal score (10 pts), health
`DCGM_FI_DEV_POWER_USAGE`	power_w	Per-GPU power limit filter
`DCGM_FI_DEV_XID_ERRORS`	xid_errors	Nonzero marks the GPU degraded

The rest of the 100-point score: 20 for meeting hard requirements, 15 for health, 10 for matching a preferred model. GPUs that are reserved, unhealthy, short on free VRAM or over the power limit are filtered out first. Multi-GPU jobs are placed on a single node.

API

Endpoint	Purpose
`GET /health`	Liveness check
`GET /inventory`	Normalized GPU inventory
`POST /recommend`	Best node, GPU ids, score, reasons and warnings. Returns 404 with per-GPU rejection reasons if nothing fits.
`POST /plan`	Execution plan steps (not executed in v0.1)

Configuration

Variable	Meaning
`SUPER_INU_METRICS_URL`	dcgm-exporter endpoint, e.g. http://host:9400/metrics
`SUPER_INU_METRICS_FILE`	Saved or mock Prometheus text file
`SUPER_INU_PROVIDER_ID`, `SUPER_INU_NODE_ID`	Provider label, and node fallback when metrics have no Hostname

Limitations

Recommends only. Launching and releasing real workloads arrives in v0.2.
No pricing yet, so budgets are recorded but not enforced.
Single-node placement; weights are untuned.
Tested on mock data, not yet on live hardware.

Own the decision, not the metal.

Agents are about to become buyers

Explainable placement, shipping today

Every decision makes the next one better

A market that routes itself

Multi-vendor by design

Routing first, payments last

Today

With Super Inu

Advanced compute routing modules

Telemetry Reader

Inventory

Placement Engine

Explainer

Provider Interface

Execution Planner

Watch the router think.

The signals behind every decision.

GPU utilization

Framebuffer memory

Temperature

Power draw

XID errors

Tensor-core activity

NVLink bandwidth

ECC errors

Clocks and throttling

Route across clusters, clouds and marketplaces

Local cluster

Cloud A

Cloud B

GPU marketplaces

The routing flywheel

Workload

The path from router to compute economy.

Super Inu docs

Quickstart

What it reads from NVIDIA DCGM

API

Configuration

Built on NVIDIA's open-source GPU stack

NVIDIA/DCGM

NVIDIA/dcgm-exporter

NVIDIA/gpu-operator

Limitations