GPU requirements
Most competitive subnets on Bittensor today are inference subnets running large language models or vision models. These workloads are VRAM-bound: you need enough GPU memory to load the model in full at a competitive precision. For LLM inference subnets, a 24 GB VRAM GPU (such as an RTX 3090, 4090, or A5000) is the practical minimum for mid-size models, while subnets running 70B+ parameter models require 80 GB or multi-GPU setups.
GPU compute throughput (TFLOPS) matters as much as VRAM for latency-sensitive subnets. A validator that measures response time will penalize a slow GPU even if it produces correct answers — see how validators score miners. NVIDIA GPUs with high tensor core throughput (A100, H100, RTX 4090) consistently outperform older architectures on inference benchmarks. AMD GPU support varies by subnet — check the subnet repository for compatibility before purchasing.
CPU and memory
The CPU handles the Bittensor miner process, network communication, and any pre- or post-processing around your model. For most subnets, a modern 8-core CPU is sufficient — the GPU is the bottleneck, not the CPU. However, data processing subnets that run CPU-heavy pipelines may require 16 or more cores to maintain competitive throughput.
System RAM requirements are typically modest relative to VRAM: 32 GB is adequate for most inference workloads, though subnets with large dataset loading or preprocessing pipelines may benefit from 64 GB or more. Storage is less critical — an NVMe SSD with at least 512 GB free ensures fast model loading at startup without becoming a bottleneck during operation.
Software stack
Bittensor mining runs on Python 3.10+ and the bittensor SDK (installed via pip install bittensor). The SDK handles all protocol communication: registering your hotkey on the subnet, receiving queries from validators, and submitting responses. On top of the SDK, each subnet installs its own dependencies — typically a deep learning framework (PyTorch or JAX), a model serving library (vLLM, TGI, or similar), and the subnet-specific miner code from the subnet repository.
Most miners run on Ubuntu 22.04 LTS with NVIDIA drivers and CUDA installed. Windows Subsystem for Linux (WSL2) works for development and testing but is not recommended for production mining due to GPU driver overhead. Docker-based setups are common for reproducibility and easier dependency management, and many subnet repos ship a Dockerfile or docker-compose configuration ready to use.
Cost analysis
Mining costs fall into two categories: capital expenditure (hardware purchase or cloud instance reservation) and operating costs (electricity and network bandwidth). A high-end consumer GPU setup (RTX 4090) draws roughly 350–450W under load; at $0.10–0.15/kWh, that is approximately $1–2 per day in electricity per GPU. Cloud GPU instances (AWS, RunPod, Vast.ai) remove the hardware risk but add ongoing rental cost — typically $0.50–3.00/hour depending on GPU tier.
Hardware amortization matters for owned equipment. Spread the purchase cost over a reasonable useful life (18–36 months for consumer GPUs) to calculate a daily hardware cost. Sum this with electricity to get your total daily cost of operation, then compare it against your expected daily TAO emission at current TAO prices. A positive spread means you are mining profitably; a negative spread means you are subsidizing your participation.