Choosing a GPU for Gonka: Hardware Recommendations

Gonka requires an NVIDIA GPU with CUDA support and at least 40GB of VRAM on an MLNode. This is a strict hardware limitation: the Kimi K2.6 model with MoE architecture (22 billion active parameters out of 235 billion total) requires a significant amount of video memory to load weights and perform inference. AMD and Intel GPUs are not supported—Gonka uses the NVIDIA CUDA stack, including cuBLAS for matrix operations and cuDNN for neural network layers.

Full list of requirements:

• GPU: NVIDIA with CUDA, minimum 40GB VRAM. Cards with less capacity (RTX 4090 with 24GB, RTX 3090 with 24GB) are not suitable—they physically cannot load the model shards.
• CPU: Support for AVX instructions is mandatory—without them, inferenced will not start (SIGILL at startup).
• RAM: 64GB+ of RAM is recommended for comfortable operation, loading weights, and handling request queues.
• Disk: NVMe SSD with sufficient space—the complete set of Kimi K2.6 weights takes ~640GB, and fast loading from NVMe is critical for node cold start time.
• Internet: At least 100 Mbps of stable connection—the node receives requests from Transfer Agents and sends results to clients in real-time.
• Uptime: 24/7—skipping epochs reduces rewards, and prolonged downtime can lead to exclusion from the task pool.

Let's analyze each recommended card in detail:

NVIDIA H100 80GB — the current generation flagship and the optimal choice for Gonka. TDP (thermal design power) of 700W, costs ~$25—35K per card. Supports FP8 inference, which accelerates request processing without quality loss. NVLink allows combining multiple H100s into a cluster with high-speed inter-card communication. A full Kimi K2.6 cluster requires 8 H100 cards (8 x 80GB = 640GB total VRAM). This is the most common configuration in the Gonka network.

NVIDIA H200 141GB — the next generation with nearly double the memory capacity. Increased VRAM allows processing more requests simultaneously (larger batch size), which increases GNK earnings per unit of time. HBM3e memory bandwidth is higher than H100 — faster weight loading, faster inference. For a Kimi K2.6 cluster, 5 H200 cards are sufficient instead of 8 H100s, simplifying infrastructure.

NVIDIA A100 40/80GB — the previous generation, but still supported by the network. Price is ~$10—15K per card — significantly cheaper than H100. Performance is lower: no FP8, slower HBM2e. The 40GB version is the minimum allowed for Gonka, the 80GB version is preferable. A100 is suitable for entering the network with lower initial investments.

What is not suitable: consumer cards RTX 4090 (24GB), RTX 3090 (24GB), RTX 4080 (16GB) — insufficient VRAM to work with Kimi K2.6. Even the most powerful consumer card falls short of the 40GB minimum threshold. For a full cluster (640GB VRAM), you will need 8 H100 cards, 5 H200 cards, or 8—16 A100 80GB cards, depending on the configuration.

An MLNode in the Gonka network is a server with a GPU that performs AI inference. Node setup involves several stages, each of which is critical for stable operation and maximum GNK earnings.

Software: the main component is the inferenced CLI, which manages model loading, request processing, and communication with the blockchain. Inferenced runs inside a Docker container, which simplifies deployment and updates. A full Kimi K2.6 configuration requires 640GB of total VRAM — for example, 8 H100 cards of 80GB each. Model weights (~640GB) are loaded from NVMe SSD at node startup.

Registration: after installation, the node registers on-chain — it creates a record on the Gonka blockchain specifying its address, supported models, and specifications (VRAM, bandwidth, location). From this moment, Transfer Agents begin directing AI requests from users to the node.

Network Operation: every incoming request — a user prompt — is processed by the GPU via a neural network. The result is sent back to the client via a Transfer Agent. The Sprint Consensus accounts for every performed computation when forming a block, and the reward is distributed proportionally to the volume of work. The node can publish updated specifications in real-time — if the load increases, Transfer Agents will redirect a portion of requests to less loaded nodes. Detailed configuration instructions are in the mining guide.

If you don't have your own equipment, there are three ways to access a GPU for Gonka, each with a different balance of cost, complexity, and control:

Pools (from $1 – Ancapex, from $100 – Gonka.Top): Ancapex, Gonka.Top, GonkaPool.ai, CloudMine (Mingles) – operators rent GPUs, set up nodes, and monitor uptime. You receive GNK proportionally to your contribution, without touching technical details. The ideal path for newcomers and passive investors.

Dedicated servers (from $12,000/month): Gonka.Top offers not only pools but also fully managed dedicated servers. You get a ready-made node – the operator handles inferenced setup, 24/7 monitoring, updates, and troubleshooting. Mining goes directly to your wallet – all GNK income is yours, minus a fixed rental fee.

Bare-metal rental: Spheron provides bare-metal servers with H100/H200 that you configure yourself (users from Russia may experience payment difficulties when using Spheron through a payment processor). This is a path for technical users familiar with Linux, Docker, and CLI. Maximum control, but also maximum responsibility for setup, uptime, and updates. A detailed comparison of all providers is on the “Get GNK” page.