Subnet 4: Targon

Targon is a decentralized confidential compute marketplace on Bittensor, where miners provide hardware-attested GPU and CPU workloads secured by Confidential Computing technology.

Subnet 4: Targon

Targon (netuid 4) is a decentralized confidential AI compute marketplace operated by Manifold Inc. It uses the Targon Virtual Machine (TVM) to provide hardware-attested inference workloads secured by Confidential Computing (CC) and Protected PCIe (PPCIE) technology. Miners contribute hardware that runs AI inference inside isolated, verifiable execution environments; validators score that work; and emissions flow via an on-chain auction system.

References: manifold-inc/targon (GitHub), targon.com

How the Mechanism Works

Targon’s incentive design centers on hardware-attested confidential compute. Rather than scoring raw model outputs, validators verify that miner hardware can run workloads inside a hardware-enforced Trusted Execution Environment (TEE) — either AMD SEV-SNP (CPU) or NVIDIA Confidential Compute / PPCIE (GPU).

Miner emissions are allocated through an auction system. Each hardware category (e.g., SEV-CPU-AMD-EPYC-V4, TDX-NVCC-NVIDIA-H200) has its own auction slot with a configurable emission share and minimum cluster size. Miners bid into these slots; bids above the maximum are capped. Any emission share not allocated to active auctions is burned. The current auction state can be queried at:

GET https://tower.targon.com/api/v1/auctions

The TVM handles the full execution pipeline: secure boot, hardware root of trust, remote attestation, and memory isolation — ensuring validators can cryptographically verify that a miner’s compute is running inside a genuine TEE.

References: manifold-inc/targon (GitHub), miner docs

Participating as a Miner

Targon mining requires specialized server hardware with BIOS access. Three hardware paths are supported:

  • AMD EPYC™ 4th Gen (9xx4 Series, Genoa/Bergamo) with SEV-SNP — CPU confidential workloads. Requires Ubuntu 25.04 Server and HGX Firmware Bundle v1.7.0 or higher.
  • Intel TDX + NVIDIA Hopper GPUs (H100, H200) — GPU passthrough inside TDX confidential VMs. Two Ubuntu variants are supported (22.04 LTS and 25.04+).
  • Intel TDX + NVIDIA Blackwell GPUs — Next-generation GPU passthrough, same TDX isolation model.

Setup involves configuring the host OS and BIOS for the chosen TEE technology, then using the tvm/install binary to pull and launch a miner VM image from Targon’s infrastructure. The targon-cli tool manages the running miner (updates, logs, container status). BIOS access is strongly recommended, as hardware configuration adjustments are expected during onboarding.

References: miner docs, amd-cpus setup

Participating as a Validator

Validators run on standard CPU servers and do not require TEE-capable hardware. The recommended setup uses a latitude.sh m4.metal.medium box running Ubuntu 24.04 or 25.04. The setup flow:

  1. Configure a .env file with hotkey phrase, MongoDB credentials, and validator IP.
  2. Use the tvm/install binary (with --node-type vali-cpu) to download and install the validator VM image.
  3. Launch the VM with sudo launch_vm.sh from the downloaded folder.
  4. Initialize the validator with targon-cli vali init [path to .env].

Ongoing management (logs, updates, container listing) uses targon-cli vali subcommands. Weights are set automatically by the running validator.

References: validator docs

On-Chain Identity

Targon’s on-chain identity is registered under owner coldkey 5CXGPMnq9RCCLUEvp9G2iUuabw69TSFM155UVS1S4Zmusaxv with 256 neurons active. The registered GitHub is manifold-inc/targon and the registered website is targon.com.

References: taostats.io/subnets/4

Further Reading

Topics Subnets