Compute Horde

# ComputeHorde (Subnet 12 of Bittensor)

![ComputeHorde.svg](ComputeHorde.svg)

ComputeHorde is a specialized subnet within the [Bittensor network](https://bittensor.com) 
designed to **supercharge Bittensor with scalable and trusted GPU computing power**. 

By transforming untrusted GPUs provided by miners into trusted compute resources, 
ComputeHorde enables **validators of other subnets** to access large amounts of decentralized computing power cost-effectively, 
paving the way for Bittensor to scale beyond its current limitations to support potentially over 1,000 subnets.

## Key Features

- **Decentralized Compute for Bittensor Validators**  
  ComputeHorde aims to become the go-to decentralized source for hardware needed to validate other subnets.
  The mission is to decrease the Bittensor ecosystem's dependency on centralized services.
  The assurance of the miners' work quality is essential for the Bittensor's overall reliability.
  
- **Fair and Verified Work**  
  ComputeHorde employs mechanisms to ensure miners provide authentic compute work, fairly verified by the validators:
  - Execute tasks from validators stake-proportionally
  - Handle **organic** (external, from other subnets) tasks.
  - Match jobs to the advertised hardware (e.g., ensuring A6000 GPUs are used for tasks requiring them).
  - Prevent malicious behaviors like "weight-copying" through innovative validation mechanisms.

- **Scalable Mining with Executors**  
  Each miner in ComputeHorde can spawn multiple **executors**, performing individual compute tasks. 
  This removes the 256 miner (UID) limit and significantly scales the potentially available computing power.

- **Hardware Classes**  
  ComputeHorde introduces hardware classes to create a free market for GPU resources, balancing cost-effectiveness with performance.
  Currently, **A6000** is the supported class, with **A100** coming next.
  The end goal is to eventually support all GPU types/configurations required by validators across Bittensor subnets.

- **Developer SDK for Easy Integration**  
  A Python SDK is available to help **subnet owners and validators** run jobs on ComputeHorde. 
  See [ComputeHorde SDK README](compute_horde_sdk/README.md#readme) for installation and usage examples.

- **Collateral-Based Trust for Organic Jobs**  
  Validators can now **require miners to deposit collateral** to be eligible for organic jobs.
  This increases reliability by enabling slashing of dishonest miners, improving trust in cross-subnet compute.
  - Validators automatically begin using the feature upon deploying the
  [collateral contract](https://github.com/bactensor/collateral-contracts#recommended-validator-integration-guide-as-used-by-computehorde).
  - Miners interested in organic jobs can deposit collateral by following the
  [miner guide](https://github.com/bactensor/collateral-contracts#recommended-miner-integration-guide-as-used-by-computehorde).

- **Built-in DDoS Shield for Miners**  
  Miners can now protect themselves from denial-of-service attacks by activating the **optional DDoS shield**.
  This tool reduces vulnerability and ensures stable mining performance.
  - Just run the Docker image from the [DDoS Shield repository](https://github.com/bactensor/bt-ddos-shield#running-shield-on-server-miner-side).


## Bittensor Context

Bittensor is a decentralized network designed to ensure that AI, the most critical technology of our era, remains accessible to everyone and free from the control of centralized entities.
Each Bittensor subnet specializes in a _digital commodity_, ranging from storage and large language models to general computing. 

This is achieved by distributing $TAO tokens to incentivize:
- Subnet owners to define the most useful and reliable commodities (by designing the incentive mechanism),
- Miners to deliver high-quality and efficient services innovatively,
- Validators to reward miners based on their performance.

Bittensor's end goal is to create an unstoppable, self-sustaining ecosystem free from single-point control, enabling innovation and resilience for the entire network.
ComputeHorde adds GPU-powered validation to this ecosystem, helping other subnets operate effectively without relying on centralized cloud services.

## Using ComputeHorde to Power Other Subnets

If you're a **subnet owner or validator** looking to access scalable, decentralized GPU compute power from ComputeHorde,
check out the [ComputeHorde SDK](compute_horde_sdk/README.md#readme).

The SDK allows you to:
- **Submit jobs easily** from your own subnet validator code.
- Perform **cross-validation** to verify result integrity and protect against malicious miners.
- **Reduce infrastructure costs** by eliminating the need for physical GPUs.
- **Accelerate validation** by leveraging ComputeHorde’s ready-to-go GPU pool.
- **Fallback to other clouds** like [RunPod](https://www.runpod.io/) via SkyPilot if ComputeHorde is temporarily unavailable.

## Scoring Mechanism

The scoring mechanism in ComputeHorde is designed to **incentivize miners to perform organic jobs** 
while maintaining accountability and fairness in the network. 

### Big Picture
- **Validators earn compute allowance.** Each block mints executor-seconds per miner–validator pair (precisely, per miner–validator–executorClass). 
  The total supply is proportional to the number of executors (GPUs) in the subnet, and each validator’s share scales with stake.  
- **Fair usage is enforced.** Because allowance is minted per pair, validators must spread jobs across miners proportionally to fully use their share.  
- **Validators pay with blocks.** Starting a job consumes allowance blocks equal to its runtime.  
- **Scoring is global.** All validators score all miners’ work, regardless of who paid, as long as jobs fit valid allowance blocks.  
- **Runtime converts to points.** Paid seconds per miner are summed, folded by coldkey, then split across hotkeys 
  following the [dancing rules](#dancing-bo